A bridge safety evaluation method, device, equipment, medium and program product

Abstract

The application belongs to the technical field of safety evaluation, and particularly relates to a bridge safety evaluation method, device, equipment, medium and program product. The application solves the safety blind area problem of pure semantic routing in a high-risk physical scene, that is, the problem that even if the user language expression does not match the current physical risk state, the routing result can still be corrected to the correct emergency skill in the high-risk physical scene.

Description

Technical Field

[0001] This invention belongs to the technical field of safety assessment, and specifically relates to a method, apparatus, equipment, medium, and program product for safety assessment of bridges. Background Technology

[0002] With the continuous advancement of transportation infrastructure construction in my country, long-span bridges face significant challenges in complex service environments such as climate aging, vehicle overloading, and geological disasters. As a critical node in the transportation network, the structural safety of bridges directly affects public safety. Establishing a bridge structural health monitoring (SHM) system and utilizing sensing technology and data analysis to achieve disaster early warning and maintenance assessment is an important technical means to ensure the long-term safe operation of bridges.

[0003] Traditional bridge health monitoring systems have evolved from manual inspection-dominated systems and single-sensor threshold alarms to the current integrated management of multi-source sensor data. In engineering practice for long-span special bridges (such as suspension bridges and tied-arch bridges), existing systems typically deploy a large network of sensors at key bridge locations. These sensors include temperature and humidity meters and ultrasonic anemometers to monitor the environment, and triaxial accelerometers, hydrostatic levels, biaxial inclinometers, and surface strain gauges to monitor structural response. This underlying monitoring data is aggregated through an edge computing gateway and transmitted back to a cloud database. Pre-defined mathematical models, such as Principal Component Analysis (PCA) and autoencoder neural networks, are used to extract potential features from the data and generate safety assessment indicators.

[0004] With the rise of large language models, related fields have begun to explore their application in intelligent bridge operation and maintenance. In AI agent systems based on large language models, a "skill" refers to a functional unit that encapsulates specific business logic, such as "daily inspection report generation," "structural health assessment," and "emergency finite element inversion analysis." Each skill corresponds to a complete execution process. The system centrally stores all available skills in a skill library and selects the most suitable skill from the library for execution based on the user's input natural language commands through a skill routing mechanism. Currently, similar implementation schemes often involve inputting bridge health status data into the large model or constructing a general fine-tuning model to assist in generating inspection reports. For example, the patent CN120106123B, "A Multi-Agent Interaction Method Based on a Large Model for Intelligent Bridge Maintenance," trains and fine-tunes a large-scale bridge foundation model using multi-dimensional bridge domain datasets, and combines it with a multi-agent system for collaborative interaction and management decision-making. Such solutions rely entirely on text semantic similarity for skill routing—that is, semantically matching user commands with skill description text and selecting the most semantically similar skill to execute. However, bridge structural health monitoring involves real-time sensor data acquisition and the issuance of physical execution commands, which is fundamentally different from pure text question-and-answer scenarios—skill scheduling errors can lead to irreversible consequences in the physical system. Existing solutions have not yet addressed the safety issue of skill scheduling in high-risk physical scenarios for large models. Summary of the Invention

[0005] This application addresses the safety blind spot problem of purely semantic routing in high-risk physical scenarios—the issue of mismatch between user language descriptions and the current physical danger state—by providing a bridge safety assessment method, apparatus, equipment, medium, and program product. It achieves the technical effect of correcting routing results to the correct emergency skills even in high-risk physical scenarios. To solve the aforementioned technical problems, this application proposes five aspects.

[0006] In a first aspect, this application proposes a bridge safety assessment method, comprising: generating an assessment task for a target bridge in response to a natural language instruction input by a user, and obtaining physical feature vectors of each phase of the target bridge; determining the current hazard level of the target bridge based on the physical feature vectors of each phase; determining an execution flow for the assessment task based on the current hazard level and the natural language instruction; and performing a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution flow.

[0007] In some embodiments, determining the current hazard level of the target bridge based on the physical feature vectors of each phase includes: determining a comprehensive hazard index of the target bridge based on the physical feature vectors of each phase; and determining the current hazard level of the target bridge based on the comprehensive hazard index and a preset hazard level comparison table.

[0008] In some embodiments, determining the execution process of the assessment task based on the current hazard level and the natural language instruction includes: determining the similarity between each skill in the preset skill library and the natural language instruction based on a simplified metadata summary of each skill in the preset skill library; determining the main skills for performing the assessment task in the preset skill library based on the current hazard level and the similarity; and determining the execution flow based on a directed acyclic graph of the main skills.

[0009] In some embodiments, determining the primary skill for performing the assessment task in the preset skill library based on the current danger level and the similarity includes: determining the physical factor of each skill based on the skill priority level of each skill and the current danger level; determining the fusion route equalization of each skill based on the current danger level, the physical factor, and the similarity; and determining the skill whose fusion route equalization conforms to the preset rules as the primary skill.

[0010] In some embodiments, determining the execution flow based on the directed acyclic graph of the primary skill includes: determining a preset task level of the assessment task based on the natural language instruction; and when the preset task level does not match the current danger level, performing a pruning operation on the directed acyclic graph of the primary skill based on the current danger level to form the execution flow.

[0011] In some embodiments, the method further includes: when the current danger level is at least two levels higher than the preset task level, traversing all skills in the preset skill library; identifying skills whose skill priority level matches the current danger level as emergency skills; and inserting other skills among the emergency skills, excluding the primary skills, into the execution flow.

[0012] Secondly, this application proposes a bridge safety assessment device, comprising: a first acquisition module, configured to generate an assessment task for a target bridge in response to a natural language instruction input by a user, and acquire physical feature vectors of each phase of the target bridge; a first determination module, configured to determine the current hazard level of the target bridge based on the physical feature vectors of each phase; a second determination module, configured to determine an execution flow for the assessment task based on the current hazard level and the natural language instruction; and a first prediction module, configured to perform a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution flow.

[0013] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in any of the first aspects.

[0014] Fourthly, this application proposes a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0015] Fifthly, this application proposes a computer program product, including a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the method described in any one of the first aspects.

[0016] This application provides a bridge safety assessment method, comprising: generating an assessment task for a target bridge in response to a user-input natural language command, and obtaining physical feature vectors of each phase of the target bridge; determining the current hazard level of the target bridge based on the physical feature vectors of each phase; determining an execution flow for the assessment task based on the current hazard level and the natural language command; and performing a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution flow. This method solves the safety blind spot problem of pure semantic routing in high-risk physical scenarios—the problem of mismatch between user language description and the current physical hazard state—and achieves the technical effect of correcting the routing results to the correct emergency skills even in high-risk physical scenarios. Attached Figure Description

[0017] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0018] Figure 1 A main flowchart of a bridge safety assessment method provided in this application embodiment; Figure 2 A main structural block diagram of a bridge safety assessment device provided in this application embodiment; Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] With the continuous advancement of transportation infrastructure construction in my country, long-span bridges face significant challenges in complex service environments such as climate aging, vehicle overloading, and geological disasters. As a critical node in the transportation network, the structural safety of bridges directly affects public safety. Establishing a bridge structural health monitoring (SHM) system and utilizing sensing technology and data analysis to achieve disaster early warning and maintenance assessment is an important technical means to ensure the long-term safe operation of bridges.

[0020] Traditional bridge health monitoring systems have evolved from manual inspection-dominated systems and single-sensor threshold alarms to the current integrated management of multi-source sensor data. In engineering practice for long-span special bridges (such as suspension bridges and tied-arch bridges), existing systems typically deploy a large network of sensors at key bridge locations. These sensors include temperature and humidity meters and ultrasonic anemometers to monitor the environment, and triaxial accelerometers, hydrostatic levels, biaxial inclinometers, and surface strain gauges to monitor structural response. This underlying monitoring data is aggregated through an edge computing gateway and transmitted back to a cloud database. Pre-defined mathematical models, such as Principal Component Analysis (PCA) and autoencoder neural networks, are used to extract potential features from the data and generate safety assessment indicators.

[0021] With the rise of large language models, related fields have begun to explore their application in intelligent bridge operation and maintenance. In AI agent systems based on large language models, a "skill" refers to a functional unit that encapsulates specific business logic, such as "daily inspection report generation," "structural health assessment," and "emergency finite element inversion analysis." Each skill corresponds to a complete execution process. The system centrally stores all available skills in a skill library and selects the most suitable skill from the library for execution based on the user's input natural language commands through a skill routing mechanism. Currently, similar implementation schemes often involve inputting bridge health status data into the large model or constructing a general fine-tuning model to assist in generating inspection reports. For example, the patent CN120106123B, "A Multi-Agent Interaction Method Based on a Large Model for Intelligent Bridge Maintenance," trains and fine-tunes a large-scale bridge foundation model using multi-dimensional bridge domain datasets, and combines it with a multi-agent system for collaborative interaction and management decision-making. Such solutions rely entirely on text semantic similarity for skill routing—that is, semantically matching user commands with skill description text and selecting the most semantically similar skill to execute. However, bridge structural health monitoring involves real-time sensor data acquisition and the issuance of physical execution commands, which is fundamentally different from pure text question-and-answer scenarios—skill scheduling errors can lead to irreversible consequences in the physical system. Existing solutions have not yet addressed the safety issue of skill scheduling in high-risk physical scenarios for large models.

[0022] While the above-mentioned solutions have achieved some success in the field of bridge monitoring, existing large-scale model application solutions still have the following shortcomings when facing complex and sudden disaster events and refined maintenance scenarios: Defect 1: High-frequency time-series sensor data causes context overload in large models. The context window of a large language model is designed for low-frequency discrete text, while the dynamic sensors of long-span bridges (such as triaxial accelerometers and strain gauges) capture the natural frequencies of the structure, typically using continuous sampling at high frequencies of 50Hz to 100Hz. If high-frequency time-series data is directly translated into a text array and fed into a large model, in high-frequency, long-time sampling scenarios (for example, sampling at 100Hz for 60 seconds generates approximately 6,000 data points, which translates to approximately 30,000 to 60,000 characters of text), it usually exceeds the context window limit of the large model, resulting in excessive computational resource consumption and failure to perform effective data dimensionality reduction and feature extraction.

[0023] Defect 2: The general agent framework lacks physical boundary protection, posing a risk of irreversible control. In emergency situations involving bridge disasters, the inference logic of large models is susceptible to interference from complex environmental variables, leading to erroneous outputs. Existing general-purpose agent frameworks (such as standard function calling mechanisms) are natively designed for pure software environments, allowing calls to all tool interfaces by default, without setting independent access controls for physical execution tools. Bridge engineering is a high-risk physical system. If the control of physical equipment is directly exposed to the tool calling interfaces of general-purpose agent frameworks, once the model generates incorrect instructions, it may erroneously trigger alarm linkages (such as initiating unnecessary traffic control or dispatching maintenance personnel for high-altitude operations) or issue incorrect control commands to the execution equipment, causing irreversible physical consequences or personnel safety risks. The root cause of this deficiency lies in the lack of an independent physical adaptation layer that embeds physical rule constraints in the underlying communication links, decoupling the large model from the real physical equipment. Defect 3: The existing semantic routing mechanism cannot perceive physical dimensions, and will still select low-priority skills during a catastrophe. Existing multi-agent skill frameworks use pure text semantic similarity for skill distribution. Their routing hubs lack the ability to recognize civil engineering physical dimensions (such as displacement (mm) and fundamental frequency drift (Hz)) and use this as a basis for dynamic weight adjustment. For example, when the fundamental frequency of a bridge's main beam suddenly drifts from the normal 0.50Hz to 0.38Hz (drift rate 0.12Hz / min), and the strain exceeds the limit by 45%, the structure is already in a state of severe disaster. However, if the maintenance personnel input "Please check today's bridge status report," pure semantic routing will match this instruction with the "Daily Inspection Report" skill (semantic similarity 0.82), rather than the "Emergency Finite Element Inversion Analysis" skill (semantic similarity only 0.43), affecting the timing of disaster investigation. The correct approach should be to prioritize high-priority skills such as "Emergency Finite Element Inversion Analysis" based on the current physical hazard state, rather than relying solely on the user's verbal expression to determine skill selection.

[0024] To address the aforementioned technical problems, this invention proposes a bridge safety assessment method. The implementation details of this bridge safety assessment method are described below. The following content is for illustrative purposes only and is not essential for implementing this solution.

[0025] Example 1: like Figure 1 As shown, this application provides a bridge safety assessment method applicable to electronic production equipment, which can be a server, mobile terminal, computer, cloud platform, etc. The data processing functionality provided in this application embodiment can be implemented by the processor of the electronic production equipment calling program code, wherein the program code can be stored in a computer storage medium. The bridge safety assessment method includes: Step S1: In response to the natural language instructions input by the user, generate an evaluation task for the target bridge and obtain the phase physical feature vectors of the target bridge.

[0026] In this application, upon receiving a natural language command input from a user, an evaluation task for the target bridge is automatically generated, and various physical feature vectors of the target bridge are obtained. Each physical feature vector refers to an ordered set formed according to a preset rule. These physical features are extracted from raw sensor data. Specifically, high-frequency raw sensor data (such as a 100Hz continuous sampling stream) is automatically intercepted at the edge, and sliding window root mean square (RMS) calculation and fast Fourier transform (FFT) spectral feature extraction are performed. Up to 500 discrete data points per second (taking 100Hz sampling and concurrent operation of 5 types of sensors as an example, where 500 points represent the total of concurrent operation of 5 types of sensors, and 100 points / second for a single-channel accelerometer) are dimensionality-reconstructed into a low-dimensional JSON structure containing only key physical features (such as fundamental frequency drift rate, RMS bias, etc.), which is then reported to the upper-layer routing module. This dimensionality reduction process solves the problem of large model context overload and provides standardized physical feature inputs for subsequent operations.

[0027] Step S2: Determine the current hazard level of the target bridge based on the physical feature vectors of each phase.

[0028] Unlike existing technologies, this application does not simply do whatever the user requests. That is, although step S1 forms an evaluation task, how and to what extent this task should be completed is not entirely determined by the useful requirement instructions, but also needs to be judged in conjunction with the actual condition of the bridge. In this application, the indicator for judging the actual condition of the bridge is its current hazard level.

[0029] Therefore, in some embodiments, step S2, "determining the current hazard level of the target bridge based on the physical feature vectors of each phase," includes: Step S21: Determine the comprehensive hazard index of the target bridge based on the physical characteristic vectors of each phase.

[0030] Step S22: Determine the current hazard level of the target bridge based on the comprehensive hazard index and the preset hazard level comparison table.

[0031] For each physical feature x_i Calculate its normalized deviation relative to the historical baseline: d_i = |x_i - μ_i| / σ_i in μ_i This is the moving average of this characteristic over the past 30 days. σ_i This corresponds to the sliding standard deviation. This normalization process will take the dimensional engineering parameters ( Hz / min The unified mapping of physical quantities (such as % and %) to dimensionless bias scalars lays the mathematical foundation for the cross-space fusion of physical dimensions and semantic vectors.

[0032] The formula for calculating the overall risk index D is: D = max(d_i) + λ × Σ_{i: d_i>θ_anom} d_i in λ Multi-point joint over-limit penalty coefficient (default value) 0.3 ), θ _ anom The threshold for mild anomalies (default value) 1.0 The summation term only applies to normalization deviations exceeding [a certain threshold]. θ _ anom The formula is based on the accumulation of feature values. It considers both single-point extrema and multi-point joint exceedances, avoiding misjudgments caused by false alarms from a single sensor. It should be noted that in the above formula max(d_i) There is overlap between the term and the summation term—that is, the feature with the largest normalization bias is counted twice: once as... max The term is used as part of the summation term. This is intentionally designed to assign extra penalty weight to single-point extreme anomalies, so that extreme deviations in a single sensor can significantly increase the overall risk index D, thus helping to trigger high-level early warnings in a timely manner.

[0033] The hazard level comparison table in this application has a preset range of comprehensive hazard indexes for each hazard level. Therefore, after determining the comprehensive hazard index, the current hazard level can be easily determined through the hazard level comparison table.

[0034] As an example, the hazard level comparison table in this application is shown below:

[0035] Step S3: Determine the execution flow for the assessment task based on the current hazard level and the natural language instructions.

[0036] The biggest difference between this application and existing technologies is that this application does not simply do what the user requests, but rather determines what should be done by combining the user's requirements with the actual physical parameters. Therefore, in determining the execution flow of the assessment task, this application, in addition to referring to the natural language instructions input by the user, also needs to consider the current hazard level of the target bridge. These two factors are used together to form the execution flow of the assessment task.

[0037] In some embodiments, step S3, "determining the execution flow for the assessment task based on the current hazard level and the natural language instructions," includes: Step S31: Determine the similarity between each skill in the preset skill library and the natural language instruction based on the simplified metadata summary of each skill in the preset skill library.

[0038] Step S32: Determine the main skills for performing the assessment task from the preset skill library based on the current danger level and the similarity.

[0039] In some embodiments, step S32, "determining the primary skills for performing the assessment task from the preset skill library based on the current danger level and the similarity," includes: Step S321: Determine the physical factor of each skill based on its skill priority level and the current danger level.

[0040] Step S322: Determine the fusion route of each skill equally based on the current danger level, the physical factor, and the similarity.

[0041] Step S323: Determine the skills that conform to the preset rules for fusion routing as the main skills.

[0042] This application includes a pre-defined skill library containing various skills. The metadata for each skill k in the skill library includes a physical trigger priority field. tier_k (i.e., the skill priority level in this application), this field is pre-marked by the system administrator based on the business nature of the skill when it is added to the skill database. The value is... T0 (daily), T1 (Mild) T2 (Moderate) T3 (Emergency), corresponding to danger levels L=0 / 1 / 2 / 3 respectively (i.e. T0 L=0, T1 L=1,T2 L=2, T3 L=3, tier_k equal to L" refers to tier_k The corresponding hazard level number is the same as the current L). For example, the "Daily Inspection Report" is labeled as... T0 "Structural health assessment" is marked as T2 "Emergency finite element inversion analysis" is marked as T3 .

[0043] To avoid resource waste caused by injecting the full skill context, this application generates a concise metadata summary for each skill in the skill library. When determining the primary skill, only the concise metadata summaries of all skills (skill_id, name, desc, tier fields) are exposed to the large language model. Only after the primary skill is determined will the detailed engineering evaluation steps and execution constraints of that skill be injected into the current active session context.

[0044] The main skills in this application are those used to construct the execution flow. When constructing the execution flow, a comprehensive consideration of user needs and the current risk level is required. Specifically, when considering user needs, a large language model is first used to extract user intent vectors from the natural language instructions input by the user. Then, based on the user intent vectors and a simplified metadata summary of each skill, the cosine similarity between each skill and the natural language instructions is calculated.

[0045] Then, the combination weight is determined based on the current danger level when combining with cosine similarity. The fusion route score is then calculated based on the combination weight, cosine similarity, and the physical factors of each skill. The physical factors are... P_physical(k) The calculation rules are as follows: like tier_k The corresponding hazard level is equal to the current hazard level L: P_physical(k) = + γ × D (Incentive Bonus Points); like tier_k Below the current danger level L: P_physical(k) = - γ × D (penalty deduction); like tier_k Higher than the current danger level L: P_physical(k) = 0 (No intervention); Among them D The overall risk index, γ This is the physical intervention intensity coefficient, the default value. 0.4 .

[0046] The technical formula for calculating the fusion route score for each skill k in the skill library is as follows: W_final(k) = α(L) × S_semantic(k) + β(L) × P_physical(k) in S_semantic(k) The cosine similarity between the user intent vector and the skill k embedding vector (values ​​[0,1]). α (L) and β (L) represents a weighting coefficient that is dynamically adjusted according to the physical hazard level L. The weighting coefficient is determined based on a weighting coefficient table, which is as follows:

[0047] After determining the weighting coefficients based on the current danger level and calculating the cosine similarity and physical factors, the fusion route score for each skill corresponding to the current danger level and user needs can be determined.

[0048] This mechanism ensures semantic dominance under normal conditions (avoiding false alarms) and physical mandatory coverage under disaster conditions (ensuring security). α (L) and β (L) satisfies α (L)+ β The normalization constraint (L)=1 ensures that... W_final The relative ordering can reflect the physical security priority; when physical excitation terms β × P_physical When it is large, W_final The absolute value can exceed the range of [0,1], but the relative size relationship between skills can still be used for unified comparison and sorting.

[0049] Therefore, when determining the primary skills, all candidate skills are categorized as follows: W_final Sort the skills in descending order and select the top K skills (default K=3) as the candidate set. If the score of the first-ranked skill exceeds the direct adjustment threshold for a single skill... θ _ trigger ( θ _ trigger default value 0.75 This threshold is set based on empirical statistics under normal semantic dominance; when the physical excitation term β × P_physical When it is large, W_ final It may far exceed this threshold, which will also trigger a direct single-skill call; otherwise, it will enter a multi-skill collaborative orchestration process, where the multi-skill collaborative orchestration process refers to... Top -K skills are coordinated and arranged, at which point the main skill becomes the front skill. Top The skill is -K, while the main skill when directly calling a single skill is the skill that ranks first in the fusion route score.

[0050] like W_final A negative value indicates that the skill's physical priority is lower than the current danger level, and the physical penalty exceeds the semantic similarity contribution; the system will still treat it as such. W_final sort in descending order Top -K, negative skill values ​​rank lower; if Top - All skills in K W_final If all values ​​are negative, the system enters the conflict and circuit breaker determination process. Conflict circuit breaker rule: The system completes... Top After filtering with -K, check if the following three conditions are met; if all three are met, discard the filter. Top -K filters results, triggers a circuit breaker, and forces the highest-level call to be executed. tier Level of emergency skills, ignoring semantic routing score (Note: Top In scenarios where K is all negative, D usually exceeds the severe threshold and Top - 1 tier If the value is below L, conditions 1 and 2 are satisfied, and then the three-condition joint judgment is initiated. - Condition 1: D ≥ 5.0 (The overall risk index exceeds the severe threshold). - Condition 2: Top -K filter results Top -1 skill tier The level is lower than the current physical danger level L (i.e., the priority of the semantically selected skill does not match the current danger level); - Condition 3: Physical state for N consecutive sampling periods (N≥3, default N=3, per period) 500ms All are at L=3; the system uses a 60-second sliding window, each 500ms The feature values ​​and D / L calculation results are updated once by sliding. After the circuit breaker is triggered, the system records a circuit breaker log (trigger timestamp, current D value, original semantic routing result, forced coverage result) for post-event auditing and continuous model optimization.

[0051] Once the primary skills are identified, detailed engineering evaluation steps and execution constraints for those skills can be retrieved from the skill library. These engineering evaluation steps and execution processes are visualized using a directed acyclic graph (DAG). DAG ) is used to represent this.

[0052] Step S33: Determine the execution flow based on the directed acyclic graph of the main skills.

[0053] A directed acyclic graph (DAG) contains all possible execution paths. In the normal state (L=0), a complete skill DAG is maintained, meaning all execution paths will be executed.

[0054] In some embodiments, step S33 "determining the execution process according to the directed acyclic graph of the main skill" includes: Step S331: Determine the preset task level of the evaluation task according to the natural language instruction.

[0055] Step S332: When the preset task level does not match the current danger level, perform a pruning operation on the directed acyclic graph of the main skill according to the current danger level to form the execution process.

[0056] In some embodiments, step S33 "determining the execution process according to the directed acyclic graph of the main skill" further includes: Step S333: When the current danger level is at least two levels higher than the preset task level, traverse all skills in the preset skill library.

[0057] Step S334: Determine the skill whose skill priority level matches the current danger level as an emergency skill.

[0058] Step S335: Insert other skills in the emergency skills except the main skill into the execution process.

[0059] When the physical danger level rises (that is, when the current danger level in the present application does not match the task level corresponding to the natural semantic instruction input by the user, for the convenience of description in the present application, the task level corresponding to the natural language instruction is referred to as the preset task level), dynamic pruning is performed: traverse DAG all nodes in, mark nodes whose level is <L and are not in the current execution path dependency chain as "dormant", and remove them from the current execution path (nodes in the dependency chain will not go dormant even if tier <L, so as to ensure that the pre-dependencies of high-priority skills are complete); recalculate the critical path after pruning, and update the execution order. For example, when L=3, tier <L also do not go dormant, so as to ensure that the pre-dependencies of high-priority skills are complete); recalculate the critical path after pruning, and update the execution order. For example, when L=3, DAG the critical path after pruning, and update the execution order. For example, when L=3, T0 and T1 nodes go dormant if they are not in the dependency chain; T2 nodes are retained if they are in the T3 skill's dependency chain, and go dormant likewise if they are not in the dependency chain, the DAG after the pruning operation is used to generate the execution process for the evaluation task.

[0060] When L suddenly changes from a low level to a high level (for example, L=1→L=3, that is, when the current danger level in the present application is at least two levels higher than the preset task level), emergency node insertion is performed: scan all tier = T3The skill (i.e., the emergency skill in this application) is checked to see if it is already in the current execution path; if the dependency is satisfied, it is immediately... T3 The skill is inserted at the earliest execution position after all its dependent nodes; if the dependency is not satisfied, the dependent node is inserted first, then the skill is inserted later. T3 Skills. By inserting emergency skills and obtaining the execution path through pruning, an execution flow can be formed when L mutates from a low level to a high level.

[0061] Step S4: Perform a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution process.

[0062] Once the execution process that aligns with the current situation and user needs is clearly defined, the target bridge's current physical characteristic vectors can be used to conduct a safety assessment of the target bridge based on the execution process.

[0063] This application provides examples based on the following scenarios.

[0064] Scenario: The fundamental frequency of the main beam of a cable-stayed bridge suddenly changes from its normal value. 0.50Hz Drift to 0.38Hz Drift rate 0.12Hz / min Meanwhile, the strain at the critical section exceeded the limit. 45% The maintenance personnel entered the command: "Help me check today's bridge status report."

[0065] Step 0: Acquisition of raw sensor data and MCP dimensionality reduction processing The underlying accelerometer is used to 100Hz The sampling frequency continuously collects the vibration acceleration signal of the main beam, generating 100 data points per second. MCP The temporal feature dimensionality reducer of the physical adaptation layer performs the following processing on 6000 data points from the past 60 seconds: (1) Sliding window RMS calculation: The vibration RMS value per second is calculated with a window of 1 second to obtain 60 RMS samples; the average deviation of the 60 RMS samples relative to the historical baseline is 23.5%, that is, RMS_dev=23.5%; (2) FFT spectrum analysis: Perform fast Fourier transform on 60 seconds of data, extract the peak value of the main frequency, and identify the current fundamental frequency as 0.38Hz (historical baseline 0.50Hz); compare with the estimated fundamental frequency value of the previous minute (0.50Hz), and calculate the fundamental frequency drift rate f_drift = (0.50-0.38) / 1min = 0.12 Hz / min (i.e., drift of 0.12Hz per minute); (3) Strain gauge data processing: The strain gauge reading at the critical section is 145% of the design allowable value, exceeding the limit by e_exceed = 45%; (4) Other sensor data acquisition: The magnetic flux sensor reads the cable force of the stay cable and calculates the monthly average cable force decay rate T_decay = 0.8% / month over the past 30 days; the hydrostatic level reads the vertical displacement of the main beam and calculates the excess of the design allowable value d_exceed = 12.3%; (5) Output a standardized JSON structure: {"f_drift": 0.12, "RMS_dev": 23.5, "e_exceed": 45.0, "T_decay": 0.8, "d_exceed": 12.3} (f_drift unit Hz / min), and report it to the routing layer.

[0066] Step 1: Normalization of physical characteristic quantities and determination of hazard level right Step The normalized bias is calculated for each of the five physical characteristic quantities output by 0. d_i : f_drift = 0.12 Hz / min, historical mean μ = 0.01, standard deviation σ = 0.02 d_1 = |0.12 - 0.01| / 0.02 = 5.5 e_exceed = 45%, historical mean μ = 5, standard deviation σ = 8 d_2 = |45 - 5| / 8 = 5.0 RMS_dev = 23.5%, historical mean μ = 5, standard deviation σ = 6 d_3 = |23.5 - 5| / 6 = 3.08 T_decay = 0.8% / month, historical mean μ = 0.5, standard deviation σ = 0.3 d_4 = |0.8 - 0.5| / 0.3 = 1.0 (equals θ_anom=1.0, not strictly greater than, and not included in the summation) d_exceed = 12.3%, historical mean μ = 3, standard deviation σ = 5 d_5 = |12.3 - 3| / 5 = 1.86 Exceed θ _ anom (= 1.0 ) items: d_1 = 5.5 , d_2 = 5.0 , d_3 = 3.08 , d_5 = 1.86 (4 items in total) d_4 (Not participating) D = max(5.5, 5.0, 3.08, 1.0, 1.86) + 0.3 × (5.5 + 5.0 + 3.08 + 1.86) = 5.5 + 0.3 × 15.44 = 5.5 + 4.63 = 10.13 → L = 3 (Severe catastrophe, D ≥ 5.0) Note: The summation term includes the maximum value. d_1 = 5.5 ,This is 4.2The intentional design described in Section 3—assigning double penalty weights to single-point extreme anomalies.

[0067] Step 2: Determining Adaptive Weighting Coefficients check β Dynamic coefficient table: L= 3 → α (L) = 0.20 , β (L) = 0.80 Step 3: Semantic Similarity Calculation Cosine similarity between user intent vector and each skill in the skill library (three representative skills are listed below): Daily inspection report skills ( tier = T0 ): S_semantic = 0.82 Structural health assessment skills tier = T2 ): S_semantic = 0.61 Emergency finite element inversion skills ( tier = T3 ): S_semantic = 0.43 Step 4: Calculation of physical factors Routine inspection ( T0 <L=3): P_physical = - 0.4 × 10.13 = - 4.05 Structural health assessment ( T2 <L=3): P_physical = - 0.4 × 10.13 = - 4.05 Emergency finite element inversion ( T3 = L=3): P_physical = + 0.4 × 10.13 = + 4.05 Step 5: Converged Routing Score W_final (Daily Inspection) = 0.20 × 0.82 + 0.80 × (-4.05) = 0.164 - 3.240 = - 3.076 W_final (structural health assessment) = 0.20 × 0.61 + 0.80 × (-4.05) = 0.122 - 3.240 = -3.118 W_final (Emergency Finite Element Inversion) = 0.20 × 0.43 + 0.80 × (+4.05) = 0.086 + 3.240 = +3.326 Step 6: Routing Decision and DAG Reorganization Top -1 represents emergency finite element inversion (W=+ 3.326 ). Top -1 score + 3.326 Far exceeding the threshold for direct adjustment of a single skill θ _ trigger = 0.75 (Physical excitation term) β × P_physical = 0.80 × 4.05 = 3.24 make W_final (significantly exceeding the semantic similarity range of [0,1]), the system directly executes single skill calls without needing to enter the multi-skill collaborative orchestration process.

[0068] Check the three conditions for conflict-induced circuit breaking: Condition 1, D= 10.13 ≥ 5.0 Condition 2 is satisfied. Top -K Filtering Top -1 represents emergency finite element inversion ( tier = T3 ), T3 Not lower than the current danger level L=3 ( T3 Corresponding to L=3, tier (Equal to L, but not lower than L), this condition is not met. Condition 2 is not met; all three conditions are not met simultaneously. Therefore, the conflict circuit breaker is not triggered, and the system calls normally according to the merged routing result. Top -1 skill.

[0069] DAG Restructuring Execution Process: (1) Pruning: Mark nodes whose tier is lower than the current L (L=3) and are not in the dependency chain (T0 / T1 nodes such as daily inspection reports and data statistics summaries) as dormant and remove them from the execution path; (2) Emergency insertion: Scan the skill library for tier=T3 skills. The dependent nodes of bridge_emergency_fem are sensor_calibration and data_preprocess. Both are already in the current path. The dependency is satisfied. Insert bridge_emergency_fem into the earliest execution position after its dependent nodes. (3) Final execution path: sensor_calibration → data_preprocess → bridge_emergency_fem; (4) Record the execution log: {timestamp: "2026-02-25T10:32:15", D: 10.13, route_result: "bridge_emergency_fem(T3)", semantic_top1_pure: "daily_inspection(T0)"}.

[0070] Step 7: System Output Response The system returned the following to the operations and maintenance personnel: "Structural anomaly detected (hazard level L=3, overall hazard index D=)" 10.13 ). Main beam fundamental frequency drift rate 0.12Hz / min (Normalization bias) d_1 = 5.5 Deviation from historical average 5.5 (number of standard deviations), critical section strain exceeding limits 45% (Normalization bias) d_2 = 5.0 Displacement exceeding limits 12.3% (Normalization bias) d_5 = 1.86 An emergency finite element inversion analysis has been automatically initiated; routine inspection reports have been suspended. It is recommended that the structural engineer be notified immediately for on-site verification. Step 8: Comparison with pure semantic routing If pure semantic routing is used ( α = 1.0 , β = 0.0 ): W_final(Daily Inspection) = 1.0 × 0.82 = 0.82 ← Selected W_final (Emergency Finite Element Inversion) = 1.0 × 0.43 = 0.43 Pure semantic routing will select "Daily Inspection Report," meaning the system will still perform daily inspections even under severe structural disaster conditions, failing to respond promptly to physical hazards. This invention, through a physical state feedback-driven mechanism, changes the routing result from "Daily Inspection (W=... 0.82 ")" should be corrected to "Emergency finite element inversion (W=+)" 3.326 This achieves effective coverage of semantic intent by physical security requirements.

[0071] Compared to existing skills routing technologies that rely solely on textual semantic similarity (a single dimension), this invention incorporates the normalized bias (comprehensive hazard index D) of real-time physical sensor data into the routing weight calculation. βThe (L) coefficient adaptively adjusts with the physical hazard level L, achieving a dual-track fusion of semantic and physical routing; and in the event of a severe conflict between the two, a conflict circuit breaker mechanism is used to forcibly override the semantic result. This solves the safety blind spot problem of pure semantic routing in high-risk physical scenarios—even if the user's language description does not match the current physical hazard state, the routing result can still be corrected to the correct emergency skill.

[0072] Unlike existing technologies that use static, fixed workflows, this invention's skill execution graph carries physical awareness labels for each node. Execution paths are dynamically pruned and inserted based on real-time physical hazard levels, rather than being pre-defined. This means the same skill may be activated or dormant under different physical hazard levels, and the execution path is not pre-written but generated in real-time according to the physical state. Therefore, it is possible to promptly remove low-priority tasks and insert emergency response skills when the hazard level increases, reducing disaster response delays.

[0073] Example 2: Based on the foregoing embodiments, this application provides a bridge safety assessment device. The various modules and units included in the device can be implemented by a processor in a computer device; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0074] like Figure 2 As shown, a bridge safety assessment includes: a first acquisition module 1, a first determination module 2, a second determination module 3, and a first prediction module 4.

[0075] The first acquisition module 1 is used to generate an assessment task for the target bridge in response to a natural language command input by the user, and to acquire the physical feature vectors of each phase of the target bridge. The first determination module 2 is used to determine the current hazard level of the target bridge based on the physical feature vectors of each phase. The second determination module 3 is used to determine the execution flow for the assessment task based on the current hazard level and the natural language command. The first prediction module 4 is used to perform a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution flow.

[0076] In some embodiments, the first determining module 2 is further configured to determine the comprehensive hazard index of the target bridge based on the physical feature vectors of each phase; and to determine the current hazard level of the target bridge based on the comprehensive hazard index and a preset hazard level comparison table.

[0077] In some embodiments, the second determining module 3 is further configured to determine the similarity between each skill in the preset skill library and the natural language instruction based on the simplified metadata summary of each skill in the preset skill library; determine the main skill for performing the assessment task in the preset skill library based on the current danger level and the similarity; and determine the execution flow based on the directed acyclic graph of the main skill.

[0078] In some embodiments, the second determining module 3 is further configured to determine the physical factor of each skill based on the skill priority level of each skill and the current danger level; determine the fusion route equalization of each skill based on the current danger level, the physical factor and the similarity; and determine the skill whose fusion route equalization conforms to a preset rule as the primary skill.

[0079] In some embodiments, the second determining module 3 is further configured to determine a preset task level of the assessment task based on the natural language instruction; when the preset task level does not match the current danger level, perform a pruning operation on the directed acyclic graph of the main skill based on the current danger level to form the execution flow.

[0080] In some embodiments, the second determining module 3 is further configured to, when the current danger level is at least two levels higher than the preset task level, traverse all skills in the preset skill library; determine the skills whose skill priority level matches the current danger level as emergency skills; and insert other skills among the emergency skills, excluding the main skills, into the execution flow.

[0081] The various modules in the aforementioned bridge safety assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of the device in hardware form or independently of it, or stored in the memory of the processing device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods.

[0082] Example 3: Thirdly, this application provides an electronic device, such as Figure 3As shown, it includes: at least one processor 100; and a memory 200 communicatively connected to the at least one processor 100; wherein the memory 200 stores instructions executable by the at least one processor 100, the instructions being executed by the at least one processor 100 to enable the at least one processor 100 to perform a bridge safety assessment method according to the above embodiments.

[0083] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0084] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0085] In addition, the computer device may include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and input / output devices (e.g., keyboard, mouse, speakers, etc.).

[0086] The processor can communicate with external devices via the I / O bus through wired or wireless networks.

[0087] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product / computer program product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0088] Example 4: Fourthly, this application proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0089] Computer-readable storage media can be implemented by any type of volatile or non-volatile storage device or a combination thereof. Computer-readable storage media may include, but are not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, and computer storage media (e.g., hard disks, floppy disks, solid-state drives, removable disks, CD-ROMs, DVD-ROMs, Blu-ray discs, etc.).

[0090] Computer-readable storage media may also store at least one computer-executable program / instruction, such as computer-readable instructions. Computer-readable storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Computer-readable storage media may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, a non-transitory computer-readable storage medium may be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0091] Example 5: Fifthly, this application proposes a computer program product, including a computer program / instructions, characterized in that, when the computer program is executed by a processor, it implements the steps of the method described in any one of the first aspects.

[0092] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0094] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0096] In the embodiments provided in this disclosure, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0097] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a controller to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0098] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein.

[0099] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A method for safety assessment of bridges, characterized in that, include: In response to natural language commands input by the user, an evaluation task for the target bridge is generated, and the physical feature vectors of each phase of the target bridge are obtained. The current hazard level of the target bridge is determined based on the physical feature vectors of each phase; The execution flow for the assessment task is determined based on the current hazard level and the natural language instructions; The target bridge is subjected to a safety assessment based on the physical feature vectors of each phase and the execution process.

2. The method according to claim 1, characterized in that, Determining the current hazard level of the target bridge based on the physical feature vectors of each phase includes: The comprehensive risk index of the target bridge is determined based on the physical characteristic vectors of each phase. The current hazard level of the target bridge is determined based on the comprehensive hazard index and the preset hazard level comparison table.

3. The method according to claim 1, characterized in that, The process of determining the execution of the assessment task based on the current hazard level and the natural language instructions includes: The similarity between each skill in the preset skill library and the natural language instruction is determined based on the simplified metadata summary of each skill in the preset skill library; Based on the current danger level and the similarity, the main skills for performing the assessment task are determined from the preset skill library. The execution flow is determined based on the directed acyclic graph of the main skills.

4. The method according to claim 3, characterized in that, The step of determining the primary skills for performing the assessment task from the preset skill base based on the current danger level and the similarity includes: The physical factor of each skill is determined based on the skill priority level of each skill and the current danger level, wherein the physical factor is used to characterize the degree of fit between each skill and the current danger level; The fusion route score for each skill is determined based on the current danger level, the physical factor, and the similarity, wherein the fusion route score is used to characterize the degree of adaptation between each skill and the natural language instruction and the current danger level; The skills that conform to preset rules in the fusion routing are identified as the primary skills.

5. The method according to claim 3, characterized in that, Determining the execution flow based on the directed acyclic graph of the main skills includes: The preset task level of the evaluation task is determined based on the natural language instructions; When the preset task level does not match the current danger level, a pruning operation is performed on the directed acyclic graph of the main skills according to the current danger level to form the execution flow.

6. The method according to claim 5, characterized in that, The method further includes: When the current danger level is at least two levels higher than the preset task level, all skills in the preset skill library are traversed. Skills whose skill priority level matches the current danger level are identified as emergency skills; Other emergency skills besides the primary skill are inserted into the execution process.

7. A safety assessment device for bridges, characterized in that, include: The first acquisition module is used to generate an evaluation task for the target bridge in response to a natural language command input by the user, and to acquire the phase physical feature vectors of the target bridge. The first determining module is used to determine the current danger level of the target bridge based on the physical feature vectors of each phase; The second determining module is used to determine the execution flow of the assessment task based on the current hazard level and the natural language instructions; The first prediction module is used to perform a safety assessment of the target bridge based on the physical feature vectors of each phase and the execution process.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 6.

Images