Bridge detection monitoring fusion analysis and maintenance decision method based on ai algorithm

Abstract

The application belongs to the technical field of bridge detection and maintenance, and discloses a bridge detection monitoring fusion analysis and maintenance decision method based on an AI algorithm, which comprises the following steps: acquiring disease evaluation records and physical response sequences collected by a structure monitoring system, and establishing a dual-source feature set; constructing an associated topology between disease representation and response features, performing mapping coupling, and generating a fusion state vector; extracting a structure performance degradation trajectory based on the fusion state vector, combining component weights and position factors to perform hierarchical aggregation, obtaining a bridge overall health index and a component health index sequence; inputting the health index sequence and maintenance resource constraints into a decision reasoning framework to perform maintenance opportunity determination and measure level division, and generating a candidate maintenance scheme set; and screening and integrating the candidate maintenance scheme set according to urgency sorting and resource balance principles, and outputting a final maintenance decision instruction; thus, excessive maintenance and waste of management and maintenance resources are avoided.

Description

Technical Field

[0001] This invention relates to the field of bridge inspection and maintenance technology, and more specifically, to a bridge inspection and monitoring fusion analysis and maintenance decision-making method based on AI algorithms. Background Technology

[0002] With the continuous expansion of the existing scale of transportation infrastructure and the extension of its service life, the safety management and scientific maintenance of bridge structures have become core issues in the transportation sector. In the full life cycle management of bridges, accurately assessing the structural condition and formulating reasonable maintenance strategies are of great significance for ensuring public safety, extending service life, and optimizing resource allocation. To achieve comprehensive control over the condition of bridges, the current common approach is to combine regular manual inspections with structural health monitoring. The former focuses on the periodic inspection and evaluation of appearance defects, while the latter uses sensor networks to collect physical response data such as strain, displacement, and vibration in real time.

[0003] Specifically, manual inspection produces qualitative assessment results characterized by discrete time nodes, manifested as hierarchical classifications and descriptive judgments. In contrast, the monitoring system outputs quantitative physical parameters in a continuous time series, characterized by high sampling frequency and massive volume. The two types of data differ by orders of magnitude in time scale, lack a unified semantic connection in expression, and reflect different levels of apparent defects and intrinsic performance in evaluation dimensions. Due to the lack of an effective fusion analysis mechanism, it is difficult to establish a causal mapping between the damage phenomena detected and the abnormal features in the monitoring data, and the performance degradation trends captured by monitoring cannot be linked to the maintenance plan within the inspection cycle. This fragmented state of multi-source heterogeneous information makes maintenance decisions significantly lagging and blind, which can easily lead to missing the best maintenance window and accumulating safety hazards, or it can lead to over-maintenance and waste of limited management and maintenance resources, ultimately hindering the effective implementation of refined bridge management and preventive maintenance models.

[0004] In view of this, the present invention proposes a bridge detection and monitoring fusion analysis and maintenance decision-making method based on AI algorithms to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies and achieve the above objectives, this invention provides the following technical solution: a bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms, comprising:

[0006] Step 1: Obtain the defect assessment records generated by the regular bridge inspection and the physical response sequence collected by the structural monitoring system. Perform semantic tag conversion on the defect assessment records and perform time period feature aggregation on the physical response sequence to establish a dual-source feature set under a unified time benchmark.

[0007] Step 2: Construct the association topology between disease representation and response features on the dual-source feature set, and map and couple discrete disease levels with continuous response intervals through cross-domain semantic alignment to generate a fused state vector;

[0008] Step 3: Extract the structural performance decay trajectory based on the fused state vector, and perform hierarchical aggregation by combining component weights and location factors to obtain the overall health index and component health index sequences of the bridge.

[0009] Step 4: Input the health index sequence and maintenance resource constraints into the decision reasoning framework, determine the timing of maintenance and classify the level of measures based on the decay rate threshold and the intervention benefit ratio, and generate a set of candidate maintenance schemes;

[0010] Step 5: Select and integrate the candidate maintenance plans according to their urgency and resource balance, and output the final maintenance decision instructions.

[0011] The technical effects and advantages of the bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms of this invention are as follows:

[0012] This invention transforms discrete qualitative assessments into semantically labeled disease records using semantic tagging, converting them into disease semantic tags with a unified structure. It also aggregates time-series features from physical response sequences, converting continuous quantitative parameters into statistical features corresponding to the detection cycle, and establishes a dual-source feature set under a unified time benchmark. This solves the heterogeneity problem in time scale and expression between manual detection and structural monitoring data. Furthermore, it constructs a correlation topology between disease characterization and response features on the dual-source feature set. By analyzing the consistency between the direction of change of disease intensity values ​​and the direction of change of statistical features under the same component, it establishes cross-domain correlation edges, realizing a causal mapping between apparent diseases and intrinsic performance responses. Finally, through cross-domain semantic alignment, it maps and couples discrete disease levels with continuous response intervals to generate a fused state vector, overcoming the limitations of detection and... The problem of fragmented monitoring data is addressed; a health index sequence is obtained by extracting structural performance degradation trajectories based on fused state vectors and combining component weights and location factors for hierarchical aggregation, enabling quantitative characterization and evolution trend tracking of the overall and component structural states of the bridge; the health index sequence and maintenance resource constraints are input into a decision-making reasoning framework, and maintenance timing and measure levels are determined based on decay rate thresholds and intervention benefit ratios, ensuring that maintenance planning is effectively linked to performance degradation trends and avoiding lag and blindness in maintenance decisions; candidate maintenance schemes are screened and integrated according to urgency ranking and resource balance principles, preventing the accumulation of safety hazards due to missed optimal maintenance windows and avoiding excessive maintenance that wastes management resources, thereby effectively supporting the implementation of refined bridge management and preventive maintenance models. Attached Figure Description

[0013] Figure 1This is a schematic diagram of the bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1

[0016] Please see Figure 1 As shown in this embodiment, the bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms includes:

[0017] Step 1: Obtain the defect assessment records generated by the regular bridge inspection and the physical response sequence collected by the structural monitoring system. Perform semantic tag conversion on the defect assessment records and perform time period feature aggregation on the physical response sequence to establish a dual-source feature set under a unified time benchmark.

[0018] In the entire lifecycle management of bridges, regular manual inspections and structural health monitoring systems are the two main methods for obtaining bridge condition information. Manual inspections are typically conducted periodically according to standards such as the "Technical Condition Assessment Standard for Highway Bridges," generating defect assessment records with defect type, severity level, and spatial location as core information. Structural health monitoring systems, on the other hand, continuously collect physical response data such as strain, displacement, acceleration, and temperature through a sensor network deployed at key locations on the bridge, forming a high-frequency sampling time series. Because the defect assessment records generated by manual inspections have discrete time nodes and are presented as grade classifications and descriptive judgments, while the physical response sequences output by the monitoring system have continuous time series characteristics and high sampling frequency, the two types of data differ significantly in time scale, expression format, and evaluation dimensions. Therefore, it is necessary to standardize the two types of data separately and establish a unified time benchmark to achieve subsequent fusion analysis.

[0019] In this embodiment of the invention, the sources for obtaining the defect assessment records include, but are not limited to: previous inspection reports stored in the bridge management system, on-site records uploaded by the field inspection and data acquisition terminal, and digitized transcription results of paper inspection forms. The sources for obtaining the physical response sequence include, but are not limited to: strain data collected by fiber optic grating sensors, deflection data collected by displacement gauges, vibration data collected by accelerometers, and environmental data collected by temperature and humidity sensors.

[0020] Semantic tagging was performed on the disease assessment records.

[0021] Preferably, in some possible implementations of the embodiments of the present invention, the semantic tag conversion method for disease assessment records includes: extracting the disease type field, severity level, and spatial location description from the disease assessment record; standardizing and merging the disease type field according to a preset disease classification dictionary to obtain a standard disease type code; converting the severity level into a disease intensity value according to a preset level-value lookup table; and determining the structural component identifier and relative position coordinates to which the disease belongs based on the spatial location description; and combining and encapsulating the standard disease type code, disease intensity value, structural component identifier, and relative position coordinates to generate a disease semantic tag with a unified structure.

[0022] The disease type field in the disease assessment record is usually described using natural language. Different inspectors may describe the same disease differently. For example, descriptions such as "cracks in the bottom slab of the main beam," "longitudinal cracks at the bottom of the beam," and "cracking on the lower surface of the main beam" actually refer to the same disease phenomenon. To eliminate semantic ambiguity caused by differences in description, this embodiment pre-sets a disease classification dictionary to standardize and merge the disease type field.

[0023] The disease classification dictionary adopts a three-level hierarchical structure. The first level consists of major disease categories, including five main categories: cracks, deformation, defects, material deterioration, and connection failures. The second level consists of subcategories, such as cracks, which are further subdivided into transverse cracks, longitudinal cracks, network cracks, and diagonal cracks. The third level consists of minor disease categories, which further describe the specific morphological characteristics of the cracks. Each minor disease category corresponds to a unique standard disease type code, using a three-segment encoding format of "major category code - subcategory code - minor category code". For example, "CL-01" represents cracks - longitudinal cracks - ordinary longitudinal cracks.

[0024] The conversion of severity levels uses a preset level-value conversion table. According to industry standards, disease severity is typically divided into levels 1 to 5, where level 1 indicates no disease or minor disease, and level 5 indicates severe disease requiring immediate treatment. This embodiment converts the levels into disease intensity values ​​ranging from 0 to 100. The specific correspondence is as follows: level 1 corresponds to disease intensity values ​​of 0 to 10, level 2 to 11 to 30, level 3 to 31 to 55, level 4 to 56 to 80, and level 5 to 81 to 100. The level-value conversion table is set based on the following principle: low-level diseases have a smaller impact on structural performance, so a narrower value range is used; high-level diseases have a significant and highly variable impact on structural performance, so a wider value range is used to improve differentiation.

[0025] The spatial location description analysis involves two levels: first, identifying the structural component to which the defect belongs, such as main beams, piers, abutments, supports, expansion joints, etc.; second, determining the relative position coordinates of the defect on that component. Structural component identification uses a coding format of "span number-component type-component serial number," for example, "03-GB-02" represents the second main beam in the third span. Relative position coordinates are represented using the component's local coordinate system, with values ​​ranging from 0 to 1, representing the relative position along the component's longitudinal and transverse directions, respectively.

[0026] Through the above processing, the original unstructured defect assessment records are transformed into structured defect semantic tags. The tag format is: [standard defect type code, defect intensity value, structural component identifier, relative position coordinates, detection timestamp]. For example, the original record "On June 15, 2024, a level 3 longitudinal crack was found in the bottom slab of the second main beam" is transformed into the defect semantic tag: [CL-01,43,02-GB-01,(0.5,0.5),2024-06-15].

[0027] Time-phase feature aggregation of physical response sequences.

[0028] Preferably, in some possible implementations of the embodiments of the present invention, the method for agglomerating the time-period features of the physical response sequence includes: segmenting the physical response sequence into segments according to the time intervals corresponding to the detection cycle to obtain response data segments corresponding to each detection time point; extracting statistical features for each response data segment, including mean offset, peak dispersion, trend slope, and fluctuation amplitude; and spatially clustering the statistical features of different measurement points within the same detection cycle according to the structural component affiliation to generate component-level response agglomeration features.

[0029] Since structural health monitoring systems typically use sampling frequencies ranging from 10Hz to 100Hz, while manual inspections are usually conducted quarterly or annually, there is a significant difference in time scale. Directly using raw monitoring data for fusion analysis would place enormous pressure on storage and computation due to the massive amounts of high-frequency data, and it would be difficult to establish a correspondence with discrete inspection points. Therefore, it is necessary to perform time-period feature aggregation on the physical response sequence, compressing the high-frequency continuous data into a feature representation that matches the inspection cycle.

[0030] The response data segment is divided into segments based on two adjacent detection time points, with the monitoring data within the period preceding the detection time point being divided into a single data segment. For example, if periodic detections are conducted on March 15, 2024, June 15, 2024, and September 15, 2024, then the monitoring data from March 16, 2024 to June 15, 2024 constitutes the response data segment for the second detection period.

[0031] The specific definitions of the statistical features extracted for each response data segment are as follows:

[0032] The mean offset characterizes the degree of deviation of the average physical response level within the current period from the baseline state. In this embodiment, the mean response value of the first testing period after the bridge is put into operation is used as the baseline value, and the calculation formula is as follows: In the formula, For the first The mean offset of each detection period; For the first The first detection cycle response data segment Values ​​of each sampling point; For the first The total number of sampling points in the response data segment of each detection cycle; The average response over the baseline period.

[0033] Peak dispersion characterizes the degree of dispersion of extreme values ​​in a physical response and is used to identify abrupt changes in response caused by abnormal impacts or localized damage. The calculation method is as follows: first, extract peak points in the response data segment that exceed the mean plus or minus two standard deviations; then, calculate the standard deviation of these peak points as the peak dispersion. A large peak dispersion indicates irregular extreme value fluctuations in the response sequence, which may predict localized structural damage.

[0034] The trend slope characterizes the direction and speed of the physical response's evolution within the detection period. A least-squares method is used to linearly fit the response data segment; the slope of the fitted line is the trend slope. A positive slope indicates an increasing trend in the response value, while a negative slope indicates a decreasing trend. The absolute value of the slope reflects the rate of change. The fluctuation amplitude characterizes the oscillation intensity of the physical response and is measured by the ratio of the standard deviation to the mean of the response data segment.

[0035] Since multiple measuring points are typically deployed on the same structural component, it is necessary to spatially cluster the statistical characteristics of each measuring point within the same structural component to form component-level response aggregation characteristics. Spatial clustering adopts a grouping and aggregation strategy based on component identifiers: first, the structural component identifier to which the measuring point belongs is determined according to its installation location; then, the mean and standard deviation of each statistical characteristic of all measuring points within the same structural component are calculated, with the mean representing the overall response level of the component and the standard deviation representing the response variability within the component.

[0036] The format of the component-level response cohesion characteristics is: [Structural component identifier, detection cycle number, {mean value of mean offset, standard deviation of mean offset, mean value of peak dispersion, standard deviation of peak dispersion, mean value of trend slope, standard deviation of trend slope, mean value of fluctuation amplitude, standard deviation of fluctuation amplitude}].

[0037] For example, the response cohesion characteristics of a main beam in the 5th testing cycle can be expressed as: [02-GB-01,5,{12.3,2.1,0.85,0.12,0.023,0.005,0.15,0.03}].

[0038] Establish a dual-source feature set under a unified time reference:

[0039] Disease semantic tags and component-level response cohesive features are paired and associated based on time reference and component identifier to form a dual-source feature set. The specific method of pairing and association is as follows: using the structural component identifier and the detection timestamp (or detection cycle number) as the joint key, the disease semantic tags and response cohesive features are left-outer joined. When a structural component has only disease records and no monitoring data in a specific detection cycle, the response cohesive feature field is filled with a null value marker; when there is only monitoring data and no disease records, the disease intensity value in the disease semantic tag is set to 0 (indicating that no disease was detected).

[0040] The data structure of the dual-source feature set is: {(structural component identifier, detection cycle number):[disease semantic label, component-level response agglomeration feature]}. The dual-source feature set establishes a spatiotemporal correspondence between detection data and monitoring data, laying the data foundation for subsequent correlation analysis and fusion processing.

[0041] Step 2: Construct the association topology between disease representation and response features on the dual-source feature set, and map and couple discrete disease levels with continuous response intervals through cross-domain semantic alignment to generate a fused state vector.

[0042] Because damage assessment records reflect the apparent state of structural damage, while monitoring response data reflects the intrinsic mechanical properties of the structure, they characterize the health status of bridges from different dimensions. Relying solely on a single data source for condition assessment can easily lead to biases: apparent inspections may miss hidden damage, and monitoring data may fail to explain the causes of abnormal responses. Therefore, it is necessary to establish an intrinsic correlation between the two types of data to achieve collaborative analysis of apparent damage and intrinsic responses.

[0043] Construct a correlation topology between disease characteristics and response features.

[0044] Preferably, in some possible implementations of the embodiments of the present invention, the method for constructing the association topology includes: establishing a disease node set and a response node set respectively, using the disease type code in the disease semantic tag as a node and each statistical feature quantity in the component-level response agglomeration feature as a node; calculating the co-occurrence strength between the disease node set and the response node set; when the direction of change of disease intensity value under the same component identifier is consistent with the direction of change of statistical feature quantity, establishing a positive association edge between the corresponding nodes; otherwise, establishing a negative association edge; assigning association weights to each pair of nodes according to the consistency of the number and direction of the association edges, thereby forming a cross-domain association topology graph.

[0045] The association topology graph adopts a bipartite graph structure, containing two types of nodes and edges connecting the two types of nodes. Each node in the disease node set corresponds to a standard disease type code, and each node in the response node set corresponds to a statistical feature type (mean offset, peak dispersion, trend slope, fluctuation amplitude). The edge set E connects disease nodes and response nodes, and the attributes of the edges include the association direction (positive or negative) and the association weight.

[0046] The co-occurrence intensity is calculated based on historical data from the dual-source feature set. For each pair of disease nodes and response nodes, the records of all structural components in each detection cycle are traversed, and the number of instances where significant changes in disease type and response characteristics occur simultaneously is counted. The criteria for determining significant changes are: a change in disease intensity value exceeding 10 units, or a change in statistical characteristic quantity exceeding 20% ​​of its historical average.

[0047] The rules for determining the direction of association are as follows: if in most common real-world examples, the statistical characteristic also increases when the disease intensity value increases (or the statistical characteristic also decreases when the disease intensity value decreases), then it is determined to be a positive association; conversely, if the statistical characteristic decreases when the disease intensity value increases (or the statistical characteristic increases when the disease intensity value decreases), then it is determined to be a negative association.

[0048] The calculation of association weights comprehensively considers both the number of co-existing instances and directional consistency. Let the total number of co-existing instances be... The number of instances with the same direction is Then the association weight The calculation formula is: In the formula, Indicates the proportion of directional consistency; The logarithmic weight represents the co-occurrence frequency, used to balance the impact of high-frequency co-occurrence with low-frequency co-occurrence.

[0049] The construction of cross-domain association topology graphs enables the detection of disease phenomena to establish quantitative correlations with response features in monitoring data, providing a structured knowledge representation for subsequent causal analysis and state fusion.

[0050] By mapping and coupling discrete disease levels with continuous response intervals through cross-domain semantic alignment, a fused state vector is generated.

[0051] Preferably, in some possible implementations of the embodiments of the present invention, the method for generating the fusion state vector includes: selecting node pairs with association weights exceeding a preset association threshold on the cross-domain association topology graph, and establishing a bidirectional mapping table between the corresponding disease intensity value interval and the statistical feature quantity interval; for each structural component, extracting the disease intensity value of its current disease semantic label, finding the corresponding expected interval of response features through the bidirectional mapping table, and calculating the deviation with the actual statistical feature quantity to obtain the mapping deviation value; and concatenating the disease intensity value, each statistical feature quantity, and the mapping deviation value in a preset dimension order to generate the fusion state vector of the structural component.

[0052] The preset association threshold is set based on the following criteria: the association weight should be large enough to ensure the statistical significance of the association, while retaining a sufficient number of valid association pairs to support the mapping analysis. In this embodiment, the preset association threshold is set to 1.5, and this value can be adjusted between 1.0 and 2.5 depending on the abundance of historical data. When historical data is scarce, the threshold is appropriately lowered to increase the number of usable association pairs; when historical data is abundant, the threshold is increased to ensure the quality of the associations.

[0053] The bidirectional mapping table is constructed as follows: For each pair of nodes with association weights exceeding a threshold, the distribution relationship between disease intensity values ​​and statistical features in all co-existing instances is statistically analyzed. Using a quantile partitioning method, disease intensity values ​​are divided into several intervals (e.g., 0 to 20, 21 to 40, 41 to 60, 61 to 80, 81 to 100). The mean and standard deviation of statistical features within each disease intensity value interval are calculated, establishing a mapping relationship from disease intensity value intervals to response feature value intervals.

[0054] The mapping deviation value is calculated to quantify the degree of difference between the actual monitoring response and the expected response based on the severity of the damage. For structural components during the inspection cycle... The disease intensity value is used to find the corresponding expected mean of the response characteristics through a two-way mapping table. and expected standard deviation , with actual statistical characteristics Perform deviation calculation: In the formula, For standardized mapping bias values; when A value greater than 0 indicates that the actual response is higher than expected, potentially indicating hidden damage; when... When the value is less than or equal to 0, it indicates that the actual response is lower than expected, and the impact of the defects on the structural performance may be overestimated.

[0055] The dimensional structure of the fused state vector is: [Disease intensity value, mean offset, peak dispersion, trend slope, fluctuation amplitude, mean offset mapping deviation, peak dispersion mapping deviation, trend slope mapping deviation, fluctuation amplitude mapping deviation]. The fused state vector encodes apparent disease information and intrinsic response information in the same vector space, realizing deep fusion of detection data and monitoring data. The introduction of mapping deviation values ​​enables the two types of data to mutually verify and complement each other.

[0056] Step 3: Extract the structural performance decay trajectory based on the fused state vector, and perform hierarchical aggregation by combining component weights and location factors to obtain the overall health index and component health index sequences of the bridge.

[0057] The health status of a bridge structure is a dynamic process that evolves continuously over time, and a single point-in-time assessment cannot fully reflect the evolution of the structure's performance. By extracting the characteristics of the changes in the fused state vector over time, the degradation trend of structural performance can be identified, providing a forward-looking basis for preventive maintenance.

[0058] Extracting structural performance decay trajectories based on fused state vectors.

[0059] Preferably, in some possible implementations of the embodiments of the present invention, the method for extracting the structural performance decay trajectory includes: arranging the fused state vectors of the same structural component in multiple consecutive detection cycles in chronological order to form a component state evolution sequence; using an AI algorithm to identify trends in the component state evolution sequence, extracting directional features and rate features of the state evolution, where the directional features characterize the trend direction of performance improvement or degradation, and the rate features characterize the magnitude of change per unit time; and combining and encoding the directional features and rate features to generate a decay trajectory description of the structural component.

[0060] The component state evolution sequence is a time series composed of fused state vectors. Let the structural component be... From the first testing cycle to the... The fusion state vectors for each detection period are respectively Then the component state evolution sequence can be represented as a matrix. .

[0061] This embodiment employs a Long Short-Term Memory (LSTM) network as the AI ​​algorithm for trend recognition. LSTM networks can effectively capture long-term dependencies in time series and are suitable for analyzing the state evolution of bridges with periodic and trend characteristics. The network structure includes an input layer (9 neurons, corresponding to the 9 dimensions of the fused state vector), two LSTM hidden layers (64 hidden units per layer), a fully connected layer, and an output layer. The output layer adopts a dual-branch structure: one branch outputs directional features (-1 indicates degradation, 0 indicates stability, and 1 indicates improvement), and the other branch outputs rate features (continuous values ​​representing the amount of state change in each detection cycle).

[0062] The determination of directional features is based on the comprehensive change trend of each dimension of the fused state vector. Since an increase in the value of some dimensions of the fused state vector (such as disease intensity value and mean offset) indicates state deterioration, the relationship between the direction of change of some dimension values ​​and state change needs to be determined according to the specific physical meaning. In this embodiment, the mapping relationship between changes in each dimension and performance degradation is automatically learned during the training of the AI ​​algorithm.

[0063] The rate characteristic is normalized by dividing the original change by the detection period interval (in months) to obtain the monthly decay rate. For example, if the detection period interval is 3 months and the state change of a certain component is 6 units, then the monthly decay rate is 2.0 units / month.

[0064] The decay trajectory is described using a triplet format: [directional characteristics, average decay rate, rate change trend]. The rate change trend reflects whether decay is accelerating, and its value is either accelerating, constant, or decelerating. The method for determining the rate change trend is as follows: calculate the second-order difference of the decay rate sequence for the most recent three detection periods. If the mean of the second-order difference is greater than 0.5, it is determined to be accelerating; if the mean of the second-order difference is less than -0.5, it is determined to be decelerating; otherwise, it is determined to be constant.

[0065] By combining component weights and location factors, hierarchical aggregation is performed to obtain the overall health index and component health index sequences of the bridge.

[0066] Preferably, in some possible implementations of the embodiments of the present invention, the method for obtaining the health index includes: assigning component weights according to the functional importance of structural components in the bridge load-bearing system, and assigning location factors according to the correlation between the spatial location of the structural components and traffic flow; normalizing and projecting the fused state vector of each structural component to obtain a component health baseline score, and superimposing and correcting the component health baseline score with the directional features in the decay trajectory description to obtain a sub-component health index; and weighting and summing all sub-component health indices according to component weights and location factors to obtain the overall bridge health index.

[0067] The weights of structural components are assigned based on their functional importance within the bridge's load-bearing system. This embodiment uses a combination of expert scoring and the analytic hierarchy process (AHP) to determine the component weights. Bridge structural components are categorized into three classes according to their load-bearing importance:

[0068] (1) Key load-bearing components: including main beams, piers and abutments, which bear the main load transfer function of the bridge. The component weight is set to 0.40 to 0.50.

[0069] (2) Secondary load-bearing components: including crossbeams, cap beams, and tie beams, which assist in the distribution and transfer of loads. The component weight is set to 0.20 to 0.30.

[0070] (3) Ancillary components: including bearings, expansion joints, bridge deck paving, and protective facilities. They have a small impact on load-bearing safety but affect the use function. The weight of the components is set to 0.10 to 0.20.

[0071] The sum of the weights of all components should be 1.0, with the specific values ​​adjusted according to the bridge structure type. For example, for a simply supported beam bridge, the weight of the main beam can be set to 0.45, the weight of the pier to 0.25, the weight of the bearing to 0.15, and the remaining 0.15 weight shared by other components.

[0072] The location factor is assigned based on the correlation between the spatial location of structural components and traffic flow. For structural components in high traffic flow areas (such as directly beneath main lanes), the location factor is set to 1.2 to 1.5; for components in general areas, the location factor is set to 1.0; and for components in low traffic flow areas (such as beneath sidewalks or pedestrian crossings), the location factor is set to 0.7 to 0.9. The introduction of the location factor ensures that the health status of components in high-traffic-pressure locations receives greater attention in the overall assessment.

[0073] The component health baseline score is obtained using a normalized projection method. Each dimension of the fused state vector is adjusted according to its contribution to the health status (negative values ​​are taken for disease intensity, absolute value of mapping deviation, etc.). Then, a cosine similarity is calculated with a preset health status reference vector, and the similarity value is linearly mapped to the 0-100 score range to obtain the component health baseline score. A higher score indicates a better health status.

[0074] The calculation of the component health index comprehensively considers both the component's baseline health score and the directional characteristics of its decay trajectory. The corrected formula is: In the formula, For structural components The health index of each component; As a baseline score for component health; Directional feature (-1, 0, or 1); This is the direction correction factor, taken as an empirical value of 5. This correction results in bonus points for improving components and deduction points for deteriorating components.

[0075] The formula for calculating the overall health index of a bridge is as follows: In the formula, Assess the overall health index of the bridge; For structural components Component weights; For structural components Position factor; This represents the total number of structural components.

[0076] The overall bridge health index and the health indices of each component calculated in each inspection cycle are arranged in chronological order to form a health index sequence. The health index sequence visually demonstrates the time evolution of the bridge's health status, providing a quantitative basis for maintenance decisions.

[0077] Step 4: Input the health index sequence and maintenance resource constraints into the decision reasoning framework, determine the timing of maintenance and classify the level of measures based on the decay rate threshold and the intervention benefit ratio, and generate a set of candidate maintenance schemes.

[0078] Based on health index sequences and decay trajectory descriptions, structural components requiring maintenance intervention can be identified. However, maintenance decisions must also consider resource constraints and intervention benefits to maximize maintenance effectiveness under limited resource conditions.

[0079] Construction of decision reasoning framework and handling of resource constraints.

[0080] Preferably, in some possible implementations of the embodiments of the present invention, the decision reasoning framework adopts a hybrid architecture combining a rule engine and an optimization algorithm. The rule engine is responsible for making preliminary determinations on maintenance timing and classifying measures according to preset thresholds and rules; the optimization algorithm is responsible for screening and optimizing candidate solutions under resource constraints.

[0081] The specific structure of the decision reasoning framework includes the following modules:

[0082] (1) Input interface module: Receives health index sequence, decay trajectory description and maintenance resource constraint parameters;

[0083] (2) Rule base module: Stores decision-making knowledge such as decay rate threshold, measure-benefit comparison table, and intervention priority rules;

[0084] (3) Reasoning engine module: Performs logical reasoning on the input data based on the rule base to generate preliminary decision suggestions;

[0085] (4) Constraint processing module: Transforms resource constraints into constraints for the optimization problem;

[0086] (5) Optimization solution module: Integer programming or heuristic algorithms are used to solve constrained optimization problems;

[0087] (6) Output interface module: Outputs a set of candidate maintenance schemes and optimization results.

[0088] Maintenance resource constraints include the total annual maintenance budget, the duration of each operation, and traffic control windows. The total annual maintenance budget is determined by the bridge management unit based on government funding and maintenance planning. In this example, a medium-sized highway bridge is used as an example, with the total annual maintenance budget set at 2 million yuan.

[0089] When processing the annual maintenance budget, the decision-making reasoning framework pre-allocates the total annual budget by quarter. The pre-allocation strategy considers seasonal factors and the historical distribution of maintenance needs: in the first quarter (January-March), due to limited winter construction conditions, the allocation ratio is set at 15%, or 300,000 yuan; in the second quarter (April-June), construction conditions are good, and the allocation ratio is set at 30%, or 600,000 yuan; in the third quarter (July-September), which is the flood season, the allocation ratio is set at 20%, or 400,000 yuan; and in the fourth quarter (October-December), construction conditions are relatively good and the annual plan needs to be completed, the allocation ratio is set at 35%, or 700,000 yuan.

[0090] The duration of a single operation is usually limited according to traffic management requirements. For important highway bridges, the duration of a single closure for construction should not exceed 8 hours; for bridges where traffic diversion can be implemented, the duration of a single operation can be extended to 24 hours. The traffic control window refers to the time period during which maintenance work is permitted, which is usually during off-peak hours at night (such as 22:00 to 6:00 the next day) or specific periods before and after holidays.

[0091] When the resource consumption of a candidate maintenance plan exceeds the available budget for the corresponding quarter, the plan is marked as "pending adjustment" and will be split or postponed in subsequent steps. When the estimated operation time of a candidate plan exceeds the single operation time limit or cannot be scheduled within the traffic control window, the plan is marked as "needs to be split" or "delayed".

[0092] The timing of maintenance and the level of measures are determined based on the decay rate threshold and the intervention benefit ratio.

[0093] Preferably, in some possible implementations of the embodiments of the present invention, the method for determining the timing of maintenance and classifying the level of measures includes: extracting the decay rate of each structural component from the health index sequence, comparing the decay rate with a preset multi-level decay rate threshold to determine the decay level of each structural component; querying a preset measure-benefit comparison library according to the decay level to obtain the optional maintenance measures corresponding to each level and their expected benefit value and resource consumption value, and calculating the intervention benefit ratio; determining the recommended measure level according to the intervention benefit ratio, and generating a set of candidate maintenance schemes.

[0094] The decay rate is extracted based on the calculated monthly decay rate. This embodiment sets three decay rate thresholds:

[0095] (1) Emergency threshold: A monthly decay rate greater than 3.0 units / month indicates that the structural performance is deteriorating rapidly and immediate intervention is required;

[0096] (2) Warning threshold: A monthly decay rate between 1.0 and 3.0 units / month indicates that the structural performance is continuously deteriorating and requires planned intervention;

[0097] (3) Observation threshold: If the monthly decay rate is less than 1.0 units / month, it indicates that the structural performance is basically stable and needs to be continuously observed and tracked.

[0098] The thresholds were set based on empirical data and statistical analysis results from the bridge industry. According to domestic and international experience in bridge maintenance and management, components with a monthly decay rate exceeding 3.0 units / month may experience a decline in health index of more than 18 units within 6 months if not addressed promptly, reaching the point where major repairs are required; components with a monthly decay rate between 1.0 and 3.0 units / month are suitable for preventative maintenance; and components with a monthly decay rate less than 1.0 unit / month may temporarily not require maintenance.

[0099] The measure-benefit comparison database is a pre-built knowledge base that stores the applicable conditions, expected benefit values, and resource consumption values ​​of various maintenance measures. Maintenance measures are classified into four levels according to intervention intensity:

[0100] (1) Routine maintenance (Level I): including cleaning, drainage unblocking, painting maintenance, etc., resource consumption value of RMB10,000 to RMB50,000 per time, expected health index recovery of 3 to 8 units;

[0101] (2) Minor repair projects (Level II): including crack sealing, local repair, support adjustment, etc., with a resource consumption value of RMB 50,000 to RMB 200,000 per time, and an expected health index recovery of 8 to 15 units;

[0102] (3) Medium-scale repair projects (Level III): including structural reinforcement, component replacement, bridge deck renovation, etc., with a resource consumption value of RMB 200,000 to RMB 800,000 per time, and an expected health index recovery of 15 to 30 units;

[0103] (4) Major repair projects (Level IV): including reinforcement or replacement of the main load-bearing structure, overall performance restoration, etc., with a resource consumption value of RMB 800,000 or more per time, and an expected health index recovery of more than 30 units.

[0104] The intervention benefit ratio is the ratio of the expected recovery of health index to the value of resource consumption. The higher the intervention benefit ratio, the better the health improvement effect obtained per unit of resource input.

[0105] For structural components marked as requiring immediate intervention, maintenance measures that can effectively curb decay in the short term are selected from the measure-benefit comparison database based on the current health index and decay rate, with the measure level having the highest intervention benefit ratio being the recommended option. For structural components marked as requiring planned intervention, the optimal measure level is selected under budget constraints, taking into account both the intervention benefit ratio and resource consumption.

[0106] The generated candidate maintenance plan format is: [Structural component identifier, decay level, intervention marker, recommended measure level, expected benefit value, resource consumption value, intervention benefit ratio]. For example: [02-GB-01, warning level, planned intervention, level II, 12, 15, 0.80].

[0107] Step 5: Select and integrate the candidate maintenance plans according to their urgency and resource balance, and output the final maintenance decision instructions.

[0108] The candidate maintenance solution set usually contains multiple items to be addressed. Under the condition of limited resources, it is necessary to prioritize and integrate the solutions to ensure that limited resources are invested in the most urgent and most effective maintenance targets.

[0109] Preferably, in some possible implementations of the embodiments of the present invention, the method for screening and integrating candidate maintenance schemes includes: for each candidate scheme in the candidate maintenance scheme set, calculating an urgency score based on the decay level and component weight of its corresponding structural component, and sorting them from high to low urgency scores; including the candidate schemes one by one in the sorting order, while accumulating resource consumption, stopping the inclusion when the accumulated resource consumption is greater than or equal to the available budget amount for the first time, forming a preliminary scheme group; checking the spatial distribution of each candidate scheme in the preliminary scheme group, and integrating them when the maintenance operations of adjacent components can be combined; arranging the preliminary scheme group in a time sequence according to the executable time period range, and generating the final maintenance decision instruction.

[0110] The urgency score is calculated by considering both the decay level and the component weight. The urgency baselines corresponding to different decay levels are: 100 points for immediate intervention, 60 points for planned intervention, and 20 points for observation and follow-up. The formula for calculating the urgency score is as follows: In the formula, S_{urg} is the urgency score; S_{base} is the urgency baseline score corresponding to the decay level; The component weights of the structural components; The monthly decay rate of the structural components; The rate adjustment coefficient is taken as an empirical value of 0.1. This formula allows high-weight components and high-decay-rate components to obtain higher urgency scores.

[0111] After sorting by urgency score, candidate solutions are included one by one from highest to lowest. For each included solution, the cumulative resource consumption increases by that solution's resource consumption value. When the cumulative value first reaches or exceeds the available budget, inclusion stops and a preliminary selection group is formed. If the resource consumption value of a high-priority solution alone exceeds the remaining budget, that solution is skipped, and subsequent solutions are attempted for inclusion.

[0112] Spatial distribution checks aim to identify maintenance operations that can be combined. If two or more candidate schemes in the initial selection group have spatially adjacent structural components (such as different main beams within the same span or different cap beams on the same pier) and similar maintenance requirements (with a difference of no more than 1), these schemes can be combined into a single comprehensive maintenance project. The combined resource consumption is typically less than the sum of the resource consumption of each individual operation (saving on costs associated with repeated site visits and equipment deployment), with the savings estimated at 15% to 25%.

[0113] The timing of interventions takes into account traffic control windows and seasonal factors. Immediate interventions are prioritized for the most recent feasible time slots, while planned interventions are scheduled for quarters with favorable construction conditions. For large-scale maintenance projects requiring segmented execution, they are divided into multiple phases according to their procedures, and each phase is scheduled for execution within a different traffic control window.

[0114] The format of the final maintenance decision instruction is: [Serial number, maintenance object (structural component identifier), time window (planned execution period), measure type (maintenance measure level and specific content), budget allocation (ten thousand yuan)]. For example:

[0115] Serial Number Maintenance objects Timing window Measure type Budget allocation 1 02-GB-01 Night of the second week of April 2025 Grade II - Crack Grouting and Sealing 18 2 03-ZZ-02 Night of the third week of April 2025 Grade I - Support Cleaning and Lubrication 3 3 01-QD-01, 01-QD-02 Week 1 of May 2025 Grade II - Pier Surface Repair (Combined Operation) 25

[0116] Once the final maintenance decision instruction is output, it can be distributed to the maintenance work unit through the bridge management information system to guide the specific maintenance implementation. Simultaneously, the decision instruction is stored as a historical record in the system for subsequent maintenance effectiveness evaluation and decision model optimization.

[0117] This embodiment addresses the heterogeneity between detection and monitoring data in terms of time scale and representation by semantically labeling the defect assessment records and aggregating time-series features of the physical response sequences. It establishes a causal mapping relationship between apparent defects and intrinsic responses by constructing a cross-domain correlation topology and generating fused state vectors. Dynamic tracking and quantitative assessment of bridge conditions are achieved by extracting decay trajectories and calculating health index sequences. Scientific and reasonable maintenance decision instructions are generated through a decision reasoning framework and resource constraint optimization. This solution overcomes the technical shortcomings of the traditional bridge maintenance system where detection and monitoring data are fragmented, improves the scientific rigor and timeliness of maintenance decisions, and contributes to the effective implementation of refined bridge management and preventative maintenance models.

[0118] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms, characterized in that, include: Step 1: Obtain the defect assessment records generated by the regular bridge inspection and the physical response sequence collected by the structural monitoring system. Perform semantic tag conversion on the defect assessment records and perform time period feature aggregation on the physical response sequence to establish a dual-source feature set under a unified time benchmark. Step 2: Construct the association topology between disease representation and response features on the dual-source feature set, and map and couple discrete disease levels with continuous response intervals through cross-domain semantic alignment to generate a fused state vector; Step 3: Extract the structural performance decay trajectory based on the fused state vector, and perform hierarchical aggregation by combining component weights and location factors to obtain the overall health index and component health index sequences of the bridge. Step 4: Input the health index sequence and maintenance resource constraints into the decision reasoning framework, determine the timing of maintenance and classify the level of measures based on the decay rate threshold and the intervention benefit ratio, and generate a set of candidate maintenance schemes; Step 5: Filter and integrate the candidate maintenance plan set according to the principles of urgency and resource balance, and output the final maintenance decision instruction; The construction of the association topology between disease characteristics and response features includes: Using the disease type code in the disease semantic tag as nodes and the statistical feature quantities in the component-level response agglomeration feature as nodes, disease node sets and response node sets are established respectively; the co-occurrence strength between the disease node set and the response node set is calculated. When the direction of change of disease intensity value under the same component identifier is consistent with the direction of change of statistical feature quantity, a positive association edge is established between the corresponding nodes; otherwise, a negative association edge is established; each pair of nodes is assigned an association weight according to the consistency of the number and direction of the association edge, forming a cross-domain association topology graph; The process of mapping and coupling discrete disease levels with continuous response intervals through cross-domain semantic alignment to generate a fused state vector includes: On the cross-domain association topology graph, select node pairs whose association weight exceeds a preset association threshold, and establish a bidirectional mapping table between the corresponding disease intensity value range and the statistical feature range. For each structural component, extract the disease intensity value of its current disease semantic label, find the corresponding expected range of response features through the bidirectional mapping table, and calculate the deviation with the actual statistical feature to obtain the mapping deviation value. Concatenate the disease intensity value, each statistical feature, and the mapping deviation value in a preset dimension order to generate the fusion state vector of the structural component. The determination of maintenance timing and classification of measures based on decay rate threshold and intervention benefit ratio includes: The decay rate of each structural component is extracted from the health index sequence. The decay rate is compared with a preset multi-level decay rate threshold to determine the decay level of each structural component. Based on the decay level, a preset measure-benefit comparison library is queried to obtain the optional maintenance measures corresponding to each level, as well as their expected benefit values ​​and resource consumption values, and the intervention benefit ratio is calculated. When the decay rate exceeds the emergency threshold, it is marked as immediate intervention; when the decay rate is between the warning threshold and the emergency threshold, it is marked as planned intervention; when the decay rate is below the warning threshold, it is marked as observation and tracking. For structural components marked as immediate intervention and planned intervention, the recommended measure level is determined based on the intervention benefit ratio, and a candidate maintenance scheme set is generated.

2. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, Semantic tagging of disease assessment records, including: Extract the disease type field, severity level, and spatial location description from the disease assessment record. Standardize and merge the disease type field according to a preset disease classification dictionary to obtain a standard disease type code. Convert the severity level into a disease intensity value according to a preset level-value lookup table. Determine the structural component identifier and relative position coordinates to which the disease belongs based on the spatial location description. Combine and encapsulate the standard disease type code, disease intensity value, structural component identifier, and relative position coordinates to generate a disease semantic tag with a unified structure.

3. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, Temporal feature aggregation of physical response sequences includes: The physical response sequence is segmented according to the time interval corresponding to the detection cycle to obtain response data segments corresponding to each detection time point; statistical features are extracted for each response data segment, including mean offset, peak dispersion, trend slope, and fluctuation amplitude; the statistical features of different measuring points within the same detection cycle are spatially clustered according to the structural component affiliation to generate component-level response cluster features; the disease semantic tags and component-level response cluster features are paired and associated according to time reference and component identification to form a dual-source feature set.

4. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, Extracting structural performance decay trajectories based on fused state vectors includes: The fused state vectors of the same structural component in multiple consecutive detection cycles are arranged in chronological order to form a component state evolution sequence. The component state evolution sequence is then subjected to trend identification using an AI algorithm to extract the directional and rate features of the state evolution. The directional features represent the trend direction of performance improvement or degradation, and the rate features represent the magnitude of change per unit time. The directional and rate features are combined and encoded to generate a decay trajectory description of the structural component. The decay trajectory descriptions of all structural components are then summarized to form a set of overall structural performance decay trajectories for the bridge.

5. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, By combining component weights and location factors for hierarchical aggregation, the overall bridge health index and component health index sequences are obtained, including: Component weights are assigned based on their functional importance within the bridge's load-bearing system, and location factors are assigned based on the spatial location of the components and their correlation with traffic flow. The fused state vector of each structural component is normalized and projected to obtain a component health baseline score. This score is then superimposed and corrected with the directional features in the decay trajectory description to obtain a sub-component health index. All sub-component health indices are weighted and summed according to component weights and location factors to obtain the overall bridge health index. The overall health index and sub-component health indices for each testing cycle are arranged in chronological order to form a health index sequence.

6. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, Maintenance resource constraints include: total annual maintenance budget, single operation duration limits, and traffic control windows; The decision reasoning framework performs the following processing when generating a set of candidate maintenance schemes: The total annual maintenance budget is pre-allocated on a quarterly basis to obtain the available budget amount for each quarter; For each candidate maintenance plan in the candidate maintenance plan set, extract its resource consumption value, and calculate the total resource consumption for each quarter. When the accumulated value exceeds the available budget for the corresponding quarter, mark the candidate plan with the excess as pending adjustment. Match the single operation duration limit with the traffic control window to determine the executable time range of each candidate plan, and split or postpone the candidate plan that is not within the time range to complete the operation.

7. The bridge detection, monitoring, fusion analysis, and maintenance decision-making method based on AI algorithms according to claim 1, characterized in that, The candidate maintenance plan set is screened and integrated according to the principles of urgency and resource balance, and the final maintenance decision instruction is output, including: For each candidate maintenance scheme in the candidate maintenance scheme set, an urgency score is calculated based on the decay level and component weight of its corresponding structural component, and the schemes are sorted from highest to lowest urgency score. Candidate schemes are then included one by one in the sorted order, while simultaneously accumulating resource consumption. Inclusion stops when the accumulated resource consumption first exceeds or equals the available budget, forming a preliminary scheme group. The spatial distribution of each candidate scheme in the preliminary scheme group is examined, and when maintenance operations for adjacent components can be combined, they are integrated to reduce duplicate entry costs. The preliminary scheme group is then arranged chronologically according to the executable time period, determining the planned execution window for each candidate scheme, and generating a final maintenance decision instruction containing the maintenance object, time window, and measure type.

Images