Bridge life-cycle maintenance strategy dynamic optimization system based on genetic algorithm

Abstract

The application relates to the technical field of decision management, in particular to a bridge whole life cycle maintenance strategy dynamic optimization system based on a genetic algorithm, which comprises an initialization module, an end cost value function generation module, a dynamic rolling optimization module and a rolling decision output module; in the initial stage of the bridge life cycle, the initialization module executes genetic algorithm optimization covering the whole life cycle once to obtain optimization process data; the end cost value function generation module generates an end cost value function by using the optimization process data to estimate the expected cost of the remaining life cycle of the bridge; the dynamic rolling optimization module executes rolling optimization at the decision time in the bridge life cycle, and comprises a rolling optimization window determination unit and a rolling optimization problem construction unit; the rolling decision output module extracts only the first action instruction from the optimal action sequence as the current optimal maintenance decision for output. The application realizes real-time rolling update of the future maintenance strategy system, and avoids huge calculation burden and decision delay caused by repeated execution of complete and zero-start genetic algorithm optimization.

Description

Technical Field

[0001] This invention relates to the field of decision management technology, specifically to a dynamic optimization system for bridge life-cycle maintenance strategies based on genetic algorithms. Background Technology

[0002] In the field of infrastructure management, particularly for the maintenance and management of major engineering facilities such as bridges, developing scientific, economical, and efficient life-cycle maintenance strategies has become a core issue of concern for public management departments, operating units, and risk control agencies. Currently, many management systems employ optimization methods based on genetic algorithms to formulate a comprehensive inspection and maintenance plan covering the entire lifecycle of a bridge in its early stages of service. These methods, by setting fixed inspection intervals and maintenance trigger levels, automate the deployment of long-term management decisions and possess a certain degree of foresight and cost control capabilities.

[0003] However, practice has shown that this "static planning" model has significant limitations in the whole life cycle management of bridges. On the one hand, existing methods generally rely on Markov chain degradation models to predict the evolution of bridge health status. However, the actual operating environment is often affected by factors such as traffic load fluctuations, abnormal weather, and unforeseen events, leading to deviations between the actual condition of the facility and the prediction. When the detection results show that the bridge condition is better or worse than expected, fixed strategies cannot effectively adjust the maintenance rhythm and intensity, which may lead to wasted maintenance resources or delayed risk management, thereby affecting the overall asset operation efficiency and safety level. On the other hand, although theoretically, the genetic algorithm can be re-executed after each detection to correct subsequent strategies through dynamic recalculation, this method faces serious computational efficiency problems in execution. Traditional genetic algorithms rely on a large number of Monte Carlo simulations to evaluate life cycle costs, and each optimization must start from scratch, which greatly restricts its timeliness and feasibility in actual management processes.

[0004] Furthermore, existing methods generally lack a mechanism for inheriting historical optimization information and cannot make full use of previous optimization results to accelerate strategy updates. This makes it difficult for management systems to adapt to the requirements of modern infrastructure asset management for "real-time decision support" and "refined operation" when faced with frequent state update needs.

[0005] Therefore, how to establish a system that can efficiently and dynamically respond to actual bridge inspection data and update future maintenance strategies in real time, while avoiding the huge computational burden and decision delay caused by repeatedly executing complete genetic algorithm optimization from scratch, is an urgent problem to be solved.

[0006] To address this, a dynamic optimization system for bridge lifecycle maintenance strategies based on genetic algorithms is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic optimization system for bridge life-cycle maintenance strategies based on genetic algorithms. By establishing a model that can efficiently and dynamically respond to actual bridge inspection data, the system enables real-time rolling updates of future maintenance strategies, avoiding the huge computational burden and decision delays caused by repeatedly executing complete, zero-based genetic algorithm optimization.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A dynamic optimization system for bridge lifecycle maintenance strategies based on genetic algorithms includes:

[0010] The initialization module is used to perform a genetic algorithm optimization covering the entire life cycle of a bridge at the initial stage of its life cycle, based on a preset full life cycle model of the bridge, to obtain optimization process data.

[0011] The terminal cost value function generation module is used to generate a terminal cost value function using the optimization process data, and to estimate the expected cost of the bridge's remaining life when the bridge is in state S at a future time t.

[0012] The dynamic rolling optimization module is used to perform rolling optimization at decision points throughout the bridge's lifespan, including:

[0013] The rolling optimization window determination unit acquires the actual bridge status data at the decision time, assesses the bridge risk based on the actual bridge status data, and determines the rolling optimization window.

[0014] The rolling optimization problem construction unit combines the remaining lifetime expected cost to construct a short-term optimization problem within the rolling optimization window, and solves the short-term optimization problem through a genetic algorithm to obtain the optimal action sequence within the rolling optimization window;

[0015] The rolling decision output module is used to extract only the first action instruction from the optimal action sequence and output it as the current optimal maintenance decision.

[0016] Preferably, the preset bridge life cycle model includes: a bridge basic information unit, a cost calculation unit, a degradation modeling unit, a maintenance effect model unit, a life cycle time management unit, and a decision variable definition unit;

[0017] The bridge basic information unit is used to store and manage the bridge's design parameters, including structural type, material properties, geometric dimensions, design loads, and environmental conditions. The cost calculation unit is used to calculate various costs, including initial construction costs, periodic inspection costs, maintenance costs at all levels, failure loss costs, and corresponding discounting methods. The degradation modeling unit is used to describe the stochastic degradation process of the bridge's technical condition level. The maintenance effect modeling unit is used to define the improvement results and duration of different maintenance measures on the bridge's technical condition level. The life cycle time management unit is used to set and manage the bridge's service life time range and provide discretization processing for time nodes. The decision variable definition unit is used to define the decision variables in the genetic algorithm, including inspection time, maintenance time, and maintenance method selection.

[0018] Preferably, the process of acquiring the optimization process data includes:

[0019] The parameters of the preset bridge life cycle model are initialized, including population size, crossover rate, mutation rate, and maximum number of iterations. An initial population is generated using the bridge basic information unit and the decision variable definition unit, where each individual in the initial population represents a complete life cycle maintenance strategy. Monte Carlo simulation is performed on each individual in the initial population, using the degradation modeling unit to simulate the bridge state evolution process and the cost calculation unit to calculate the total life cycle cost. During the genetic algorithm iteration process, intermediate data at each fitness evaluation are recorded, including time t, bridge state S, and the expected life cycle cost from time t to the end of the life cycle. After multiple generations of evolution, a large amount of sample data covering different times and states is collected to form optimization process data.

[0020] Preferably, the process of generating the terminal cost value function includes:

[0021] Sample data is extracted from the optimization process data; the sample data is fitted using machine learning methods; an initial terminal cost value function is generated based on the fitted sample data, and the initial terminal cost value function is verified and calibrated to obtain the terminal cost value function. The expected cost of the remaining lifetime period at time t and state S is calculated to obtain the expected cost of the remaining lifetime.

[0022] Preferably, the process of determining the scrolling optimized window includes:

[0023] Acquire the actual bridge status data at the decision-making time, including the time information of the current decision-making time, the bridge technical condition level, the actual technical condition parameters of each bridge component, recent inspection history records, historical degradation trend information, and changes in environmental factors and load conditions; conduct a risk assessment of the bridge based on the actual bridge status data to obtain the bridge risk level.

[0024] If the bridge risk level is in the low-risk range, the optimized window length is set to [N, N+T] years; if the bridge risk level is in the medium-risk range, the optimized window length is set to [NM, N) years; if the bridge risk level is in the high-risk range, the optimized window length is set to [NP, NM) years; where 0 <N-P<N-M<N<N+T;

[0025] The time interval of the scrolling optimization window is determined based on the length of each optimized window.

[0026] Preferably, the short-run optimization problem is constructed within a rolling optimization window, taking into account the expected cost of remaining lifetime. The specific process is as follows:

[0027] The objective function of the short-run optimization problem is set as the sum of the in-window cost and the terminal cost; wherein, the in-window cost is the simulated expected cost generated by the maintenance strategy simulation within the rolling optimization window, and the terminal cost is the expected cost of the bridge's remaining life at the end of the window, calculated by calling the terminal cost value function.

[0028] Define the decision variables for the short-term optimization problem; the decision variables are represented using variable precision.

[0029] Preferably, the process of obtaining the optimal action sequence includes:

[0030] A genetic algorithm is run on the constructed short-term optimization problem to iteratively optimize through selection, crossover, and mutation operations, generating an initial population representing different action sequences within a rolling optimization window. Degenerate modeling units and Monte Carlo simulations are used to calculate the in-window cost of each individual in the initial population within the rolling optimization window. The terminal cost value function is called to calculate the terminal cost of each individual. The objective function of the short-term optimization problem is used as the fitness function to evaluate the quality of each individual in the population and select the optimal individual. When the convergence condition is met, the action sequence corresponding to the optimal individual is output, obtaining the optimal action sequence. The optimal action sequence contains a complete maintenance and inspection plan within the rolling optimization window, including inspection time, maintenance method, and maintenance time.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. This invention, through a terminal cost value function generation module, effectively solves the problem of separating long-term benefits from short-term decisions in traditional dynamic programming methods. It can quickly and accurately estimate the expected cost of a bridge over its remaining lifespan at any time t and in any state S. This allows long-term cost benefits to be incorporated into short-term decision-making during subsequent rolling optimization. It achieves a perfect integration of short-term decisions and long-term benefits in a dynamic environment, ensuring that each decision takes into account both current realities and long-term future development, significantly improving the global optimality and economic efficiency of maintenance strategies.

[0033] 2. This invention proposes a rolling optimization window adjustment mechanism based on the actual state of the bridge. The optimization window length is intelligently determined according to the bridge's risk level, achieving deep coupling between the optimization strategy and the actual bridge condition. When the bridge is in a low-risk state, the system automatically sets a longer optimization window for long-term planning to maximize cost-effectiveness. When the bridge is in a medium-to-high-risk state, the system shortens the optimization window accordingly, increasing decision-making frequency and accuracy to ensure timely response to changes in bridge condition. This state-driven dynamic adjustment mechanism better adapts to the uncertainty and nonlinear characteristics of bridge degradation, avoiding both local optima caused by excessively frequent short-term decisions and decision lag caused by excessively long planning cycles. It achieves a dynamic balance between optimization efficiency and decision accuracy, laying the foundation for subsequent short-term optimization within the rolling optimization window.

[0034] 3. This invention establishes a variable-precision decision variable encoding mechanism, employing a differentiated precision processing strategy for decision variables across different time periods. This ensures accuracy during critical decision-making periods while significantly improving the algorithm's computational efficiency, enabling complex multi-objective optimization problems to converge to high-quality solutions within an acceptable timeframe. This balance between precision and efficiency provides a feasible technical path for optimizing maintenance decisions in large-scale bridge networks. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of a dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm provided in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram illustrating the process of dynamic optimization of bridge life-cycle maintenance strategy provided in an embodiment of the present invention;

[0037] Figure 3 This is a flowchart illustrating the process of determining the scrolling optimization window according to an embodiment of the present invention. Detailed Implementation

[0038] 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.

[0039] This invention proposes a dynamic optimization system for bridge lifecycle maintenance strategies based on genetic algorithms. By establishing a system that efficiently and dynamically responds to actual bridge inspection data, it enables real-time rolling updates of the future maintenance strategy system, avoiding the enormous computational burden and decision delays caused by repeatedly executing complete, zero-based genetic algorithm optimizations. To illustrate the effectiveness of this invention in achieving real-time rolling updates of the future maintenance strategy system and avoiding computational burden and decision delays, two embodiments will be used to demonstrate its effectiveness.

[0040] Example 1:

[0041] In this embodiment of the application, the method proposed in this invention is used to dynamically optimize the life-cycle maintenance strategy of a newly built bridge. The bridge is a prestressed concrete continuous beam bridge with a lifespan of 50 years. Figure 1 The specific structural diagram of the system of the present invention includes: an initialization module, which performs a genetic algorithm optimization covering the entire life cycle based on a preset bridge life cycle model at the initial stage of the bridge's life cycle to obtain optimization process data; a terminal cost value function generation module, which uses the optimization process data to generate a terminal cost value function to estimate the expected cost of the bridge's remaining life when the bridge is in state S at a future time t; a dynamic rolling optimization module, which performs rolling optimization at decision points within the bridge's life cycle, including a rolling optimization window determination unit and a rolling optimization problem construction unit; and a rolling decision output module, which extracts only the first action instruction from the optimal action sequence and outputs it as the current optimal maintenance decision. Figure 2 This is a flowchart illustrating the dynamic optimization of bridge lifecycle maintenance strategies. The following is based on... Figure 1 and Figure 2 The following explanation is provided regarding the content:

[0042] The initialization module is used to perform a genetic algorithm optimization covering the entire life cycle of a bridge at the initial stage of its life cycle, based on a preset full life cycle model of the bridge, to obtain optimization process data.

[0043] The preset bridge life cycle model includes: a bridge basic information unit, a cost calculation unit, a degradation modeling unit, a maintenance effect model unit, a life cycle time management unit, and a decision variable definition unit;

[0044] The bridge basic information unit is used to store and manage the bridge's design parameters, including structural type, material properties, geometric dimensions, design loads, and environmental conditions. The cost calculation unit is used to calculate various costs, including initial construction costs, periodic inspection costs, maintenance costs at all levels, failure loss costs, and corresponding discounting methods. The degradation modeling unit is used to describe the stochastic degradation process of the bridge's technical condition level. The maintenance effect modeling unit is used to define the improvement results and duration of different maintenance measures on the bridge's technical condition level. The life cycle time management unit is used to set and manage the bridge's service life time range and provide discretization processing for time nodes. The decision variable definition unit is used to define the decision variables in the genetic algorithm, including inspection time, maintenance time, and maintenance method selection.

[0045] Specifically, the bridge basic information unit stores the design parameters of the prestressed concrete continuous beam bridge, including: the structural type is a prestressed concrete continuous beam, the main materials are C50 concrete and 1860MPa prestressed steel strands, the geometric dimensions are a total length of 485 meters, a width of 32.5 meters, a main beam height of 3.5 meters, a design load of Highway-I standard, and environmental conditions set as coastal area with a chloride ion corrosion environment level of C2.

[0046] The cost calculation unit sets various cost parameters based on the actual engineering conditions of the bridges: the initial construction cost is 85 million yuan; the periodic inspection cost is 20,000 yuan per inspection; maintenance costs are determined according to the bridge's technical condition level (BCI score is divided into A level 90-100 points, B level 80-89 points, C level 66-79 points, D level 50-65 points, and E level below 50 points), with the costs for repairing bridges in technical condition levels 2, 3, 4, and 5 to level 1 being 1.2 million yuan, 2.4 million yuan, 3.6 million yuan, and 4.8 million yuan respectively; failure loss costs are determined according to the failure probability based on the bridge's technical condition level, with a failure probability of 0 for level 3 and above bridges, 0.01% for level 4 bridges, and 0.5% for level 5 bridges, with a total failure loss of 40 million yuan; the discount rate is set at 5% per annum.

[0047] The degradation modeling unit adopts a bridge technical condition level degradation model based on Markov chains, and the state transition matrix P is set as follows:

[0048] The maintenance effect model unit defines the effects of different maintenance measures: minor repair can improve the bridge's technical condition level by 1 level, lasting 5-8 years; medium repair can restore the bridge to level 2, lasting 10-15 years; major repair can completely restore the bridge to level 1, lasting 18-25 years; the traffic interruption time during the maintenance period is 3 days, 15 days, and 45 days, respectively.

[0049] The life cycle time management unit discretizes the 50-year lifespan into years, and sets the inspection interval to a selectable range of 1-10 years, keeping the maintenance decision point synchronized with the inspection point.

[0050] By establishing six core units—bridge basic information unit, cost calculation unit, degradation modeling unit, maintenance effect model unit, life cycle time management unit, and decision variable definition unit—a complete and standardized modeling system for bridge life cycle management has been constructed. This modular design not only covers all modeling requirements from basic information to decision variables, ensuring that each calculation stage uses a unified data source and parameter system, but also supports bridge applications with different structural types, material properties, and environmental conditions through an independent parameter setting mechanism. This facilitates precise calibration based on actual engineering projects, significantly improving the model's versatility, consistency, and accuracy. It provides essential basic data and calculation rules for initial life cycle optimization and subsequent rolling optimization, serving as the data foundation for the effective operation of the entire system.

[0051] Furthermore, the process of acquiring the optimization process data includes:

[0052] The parameters of the preset bridge life cycle model are initialized, including population size, crossover rate, mutation rate, and maximum number of iterations. An initial population is generated using the bridge basic information unit and the decision variable definition unit, where each individual in the initial population represents a complete life cycle maintenance strategy. Monte Carlo simulation is performed on each individual in the initial population, using the degradation modeling unit to simulate the bridge state evolution process and the cost calculation unit to calculate the total life cycle cost. During the genetic algorithm iteration process, intermediate data at each fitness evaluation are recorded, including time t, bridge state S, and the expected life cycle cost from time t to the end of the life cycle. After multiple generations of evolution, a large amount of sample data covering different times and states is collected to form optimization process data.

[0053] Specifically, the parameters of the preset bridge life cycle model are initialized as follows: population size is set to 100, crossover rate is set to 0.8, mutation rate is set to 0.1, and maximum number of iterations is set to 200.

[0054] An initial population is generated using bridge basic information units and decision variable definition units. Each individual represents a complete 50-year cycle maintenance strategy, including detection time series, maintenance trigger level series, and maintenance method series.

[0055] For each individual, 1000 Monte Carlo simulations are performed to simulate the bridge's state evolution process using a degradation modeling unit: starting from the initial technical condition level 1, the state transition probability for each year is calculated according to the state transition matrix. When the bridge's technical condition level is found to reach or exceed the preset trigger level, corresponding maintenance measures are implemented. The total cost over the entire 50-year life cycle is calculated using a cost calculation unit.

[0056] During the iterative process of the genetic algorithm, intermediate data at each fitness evaluation is recorded, including time t (t = 1, 2, ..., 50 years), bridge state S (S = 1, 2, 3, 4, 5 corresponding to technical condition levels), and the expected cost from time t to the end of the life cycle, thus obtaining the remaining lifetime expected cost V(t, S). After 200 generations of evolution, approximately 20,000 sample data points covering different times and states are collected, forming the optimization process data D = {(ti, Si, Vi) | i = 1, 2, ..., 20000}.

[0057] By establishing a data acquisition mechanism for the optimization process, the goal of extracting and accumulating valuable information from the initial genetic algorithm optimization process was achieved, laying a solid data foundation for the subsequent construction of the terminal cost value function. By recording the time, state, and expected cost triplet information at each fitness evaluation, a comprehensive understanding of the cost distribution patterns under different spatiotemporal states throughout the bridge's entire life cycle was achieved. This data collection mechanism not only fully utilizes the computational investment of the initial optimization and avoids information waste, but also lays a solid data foundation for generating an accurate terminal cost function, thus providing crucial data support for the efficient operation of the entire dynamic optimization system.

[0058] Furthermore, the terminal cost value function generation module is used to generate a terminal cost value function using the optimization process data to estimate the expected cost of the bridge's remaining life when the bridge is in state S at a future time t.

[0059] The process of generating the terminal cost value function includes:

[0060] Sample data is extracted from the optimization process data; the sample data is fitted using machine learning methods; an initial terminal cost value function is generated based on the fitted sample data, and the initial terminal cost value function is verified and calibrated to obtain the terminal cost value function. The expected cost of the remaining lifetime period at time t and state S is calculated to obtain the expected cost of the remaining lifetime.

[0061] Specifically, sample data is extracted from the optimization process data D, each sample containing time t, bridge state S, and corresponding expected cost V; the sample data is fitted using the support vector regression machine learning method.

[0062] An initial terminal cost value function was generated based on the fitted sample data, and cross-validation was used to validate and calibrate the function: the dataset was divided into training and test sets in an 8:2 ratio. The prediction error on the test set was 456,000 yuan, with a relative error of 3.2%, which met the engineering accuracy requirements; the final terminal cost value function was obtained.

[0063]

[0064] Where V(t,S) is the expected cost over the remaining life; E[] is the expectation operator; T is the bridge lifespan (50 years in this embodiment); γ is the discount factor; r is the annual discount rate; C total (τ,S τ Let ) be the state at time τ and state S. τ The total cost consists of the sum of testing fees, repair fees, failure loss fees, and user fees (detour fees, time fees, etc.);

[0065] The resulting terminal cost value function can quickly calculate the expected cost over the remaining lifetime at any time t and in any state S. For example, in year 20, when the bridge is in technical condition level 3, the function predicts an expected cost of 6.8 million yuan for the remaining 30 years; in year 30, when the bridge is in technical condition level 4, the function predicts an expected cost of 5.2 million yuan for the remaining 20 years.

[0066] By employing machine learning methods to fit the optimization process data, a terminal cost value function is generated that can quickly and accurately estimate the expected cost of remaining lifetime at any time and under any state. By transforming the complex multi-stage stochastic optimization problem into a fast function lookup problem, each rolling optimization eliminates the need to re-execute a complete Monte Carlo simulation; instead, it directly calls the pre-trained value function for cost estimation. This technological transformation not only reduces the time for a single cost assessment from seconds to milliseconds but also ensures the organic combination of long-term benefits and short-term decisions during rolling optimization, avoiding global suboptimal problems caused by local optimization. This provides a core technological guarantee for achieving truly real-time dynamic optimization.

[0067] Furthermore, the dynamic rolling optimization module is used to perform rolling optimization at decision points during the bridge's lifespan, and includes: a rolling optimization window determination unit and a rolling optimization problem construction unit;

[0068] The rolling optimization window determination unit acquires the actual bridge state data at the decision moment; assesses bridge risk based on the actual bridge state data, and determines the rolling optimization window; refer to... Figure 3 The process of determining the scrolling optimized window includes:

[0069] Acquire the actual bridge status data at the decision-making time, including the time information of the current decision-making time, the bridge technical condition level, the actual technical condition parameters of each bridge component, recent inspection history records, historical degradation trend information, and changes in environmental factors and load conditions; conduct a risk assessment of the bridge based on the actual bridge status data to obtain the bridge risk level.

[0070] If the bridge risk level is in the low-risk range, the optimized window length is set to [N, N+T] years; if the bridge risk level is in the medium-risk range, the optimized window length is set to [NM, N) years; if the bridge risk level is in the high-risk range, the optimized window length is set to [NP, NM) years; where 0 <N-P<N-M<N<N+T;

[0071] The time interval of the scrolling optimization window is determined based on the length of each optimized window.

[0072] Specifically, the risk assessment is obtained by weighting and integrating the current technical condition level, degradation rate assessment, environmental corrosivity assessment, and traffic load change assessment to obtain the bridge risk level, including low risk, medium risk, and high risk;

[0073] According to the rules for determining the window length: if the bridge risk level is in the low-risk range, the optimized window length is set to [N, N+T] years (N = 15 years, T = 5 years, i.e., 15-20 years); if the bridge risk level is in the medium-risk range, the optimized window length is set to [NM, N) years (M = 10 years, i.e., 5-15 years); if the bridge risk level is in the high-risk range, the optimized window length is set to [NP, NM) years (P = 13 years, i.e., 2-5 years).

[0074] Therefore, the time interval of the scrolling optimization window is [t] k ,t k +L], where t k Where L is the current decision moment, and L is the dynamically determined optimized window length.

[0075] By establishing a risk assessment mechanism based on actual bridge condition data and a corresponding rolling optimization window dynamic adjustment strategy, a deep coupling and intelligent response between the optimization strategy and the actual bridge condition is achieved. The optimization window length is dynamically adjusted according to real-time changes in the bridge's risk level. When the bridge is in a low-risk state, a longer window is used for forward-looking planning to maximize cost-effectiveness; when the bridge is in a high-risk state, the window is shortened to increase decision-making frequency and ensure timely response. This state-driven adaptive adjustment mechanism not only effectively balances the contradiction between optimization accuracy and computational efficiency but also fully reflects the system's deep understanding of the uncertainty and nonlinear characteristics of bridge degradation. By integrating multi-dimensional information such as the bridge's physical state, historical trends, and environmental factors into a comprehensive risk index, a scientific decision boundary is provided for subsequent precise optimization, significantly improving the pertinence and effectiveness of maintenance strategies.

[0076] Furthermore, the rolling optimization problem construction unit constructs a short-term optimization problem within the rolling optimization window by combining the expected cost of remaining lifetime, and solves the short-term optimization problem through a genetic algorithm to obtain the optimal action sequence within the rolling optimization window;

[0077] The short-run optimization problem, which incorporates the expected cost of remaining lifetime into a rolling optimization window, is as follows:

[0078] The objective function of the short-run optimization problem is set as the sum of the in-window cost and the terminal cost; wherein, the in-window cost is the simulated expected cost generated by the maintenance strategy simulation within the rolling optimization window, and the terminal cost is the expected cost of the bridge's remaining life at the end of the window, calculated by calling the terminal cost value function.

[0079] Define the decision variables for the short-term optimization problem; the decision variables are represented using variable precision.

[0080] Specifically, the costs within the window include inspection costs, repair costs, failure loss costs, and user costs;

[0081] The decision variables for short-term optimization problems are represented using variable precision: monthly precision coding is used for decisions made in the first 3 years (years 9-12), quarterly precision coding is used for decisions made in the middle 4 years (years 13-16), and annual precision coding is used for decisions made in the later 3 years (years 17-19).

[0082] By constructing a short-term optimization problem with the sum of in-window cost and terminal cost as the objective function, and employing a variable-precision decision variable encoding mechanism, accurate modeling and efficient solution of the local optimization problem are achieved. In-window cost ensures accurate evaluation of the decision's effect in the current time period, while terminal cost, through invoking a pre-trained value function, enables rapid quantification of its long-term impact. The organic combination of these two ensures that each local optimization considers both short-term costs and long-term benefits. The introduction of the variable-precision encoding strategy further improves optimization efficiency. High-precision encoding is used for short-term decisions to ensure accuracy, while low-precision encoding is used for long-term decisions to reduce computational burden. This differentiated approach not only meets the engineering requirements for decision accuracy but also significantly reduces the dimensionality of the decision space, creating favorable conditions for the rapid convergence of the genetic algorithm and achieving the optimal balance between optimization quality and computational efficiency.

[0083] Furthermore, the process of obtaining the optimal action sequence includes:

[0084] A genetic algorithm is run on the constructed short-term optimization problem to iteratively optimize through selection, crossover, and mutation operations, generating an initial population representing different action sequences within a rolling optimization window. Degenerate modeling units and Monte Carlo simulations are used to calculate the in-window cost of each individual in the initial population within the rolling optimization window. The terminal cost value function is called to calculate the terminal cost of each individual. The objective function of the short-term optimization problem is used as the fitness function to evaluate the quality of each individual in the population and select the optimal individual. When the convergence condition is met, the action sequence corresponding to the optimal individual is output, obtaining the optimal action sequence. The optimal action sequence contains a complete maintenance and inspection plan within the rolling optimization window, including inspection time, maintenance method, and maintenance time.

[0085] Specifically, the convergence condition is that the optimal solution does not improve for 20 consecutive generations.

[0086] By establishing a complete genetic algorithm solution process, including population initialization, fitness evaluation, selection, crossover, and mutation operations, high-quality solutions to short-term optimization problems and reliable output of optimal action sequences are achieved. The aforementioned technological innovations are organically integrated into an executable algorithm flow. Degenerate modeling units and Monte Carlo simulations are used to calculate in-window costs, and the terminal cost value function is invoked to assess long-term impacts. The objective function of the short-term optimization problem is used as the fitness function to guide algorithm evolution, ensuring the accuracy and reliability of the optimization results. Furthermore, the output optimal action sequence includes a complete maintenance and inspection plan within the rolling optimization window, covering key decision-making elements such as inspection time, maintenance method, and maintenance duration. This provides detailed action guidelines for subsequent rolling decision outputs, achieving an effective transformation from abstract optimization models to specific engineering decisions and ensuring the practicality and operability of the entire dynamic optimization system.

[0087] Furthermore, the rolling decision output module is used to extract only the first action instruction from the optimal action sequence and output it as the current optimal maintenance decision.

[0088] Specifically, based on the rolling optimization results, the system outputs the current optimal maintenance decision as follows: continue to use the bridge, schedule the next inspection in year 18, and do not require maintenance at present; this decision will guide the actual bridge management work until the rolling optimization process is restarted at the time of the next inspection.

[0089] By establishing an organic synergy among the initialization module, the terminal cost value function generation module, the dynamic rolling optimization module, and the rolling decision output module, a fundamental shift in bridge maintenance strategy from static planning to dynamic response has been achieved. This system architecture transforms the traditional "one-time full-cycle planning" model into a "continuous learning-dynamic adjustment" model. The initialization module provides foundational data support for subsequent dynamic optimization; the terminal cost value function generation module enables rapid quantitative evaluation of long-term benefits; the dynamic rolling optimization module performs local optimization based on real-time conditions; and the rolling decision output module ensures the practicality and operability of the decisions. These four modules work together to form an adaptive closed-loop optimization system. This avoids the limitations of traditional static methods in responding to changes in actual conditions and overcomes the excessive computational burden of complete re-optimization, ultimately achieving a balance between maintenance decision accuracy, computational efficiency, and engineering practicality.

[0090] Example 2:

[0091] In Embodiment 1, the system proposed in this invention successfully achieved real-time rolling updates of the future maintenance strategy system by establishing a system capable of efficiently and dynamically responding to actual bridge inspection data, thus avoiding computational burden and decision-making delays. To further verify the effectiveness of this invention, this embodiment also dynamically optimized the full life-cycle maintenance strategy for another highway bridge.

[0092] The initialization module is used to perform a genetic algorithm optimization covering the entire life cycle of a bridge at the initial stage of its life cycle, based on a preset full life cycle model of the bridge, to obtain optimization process data.

[0093] The preset bridge life cycle model includes: a bridge basic information unit, a cost calculation unit, a degradation modeling unit, a maintenance effect model unit, a life cycle time management unit, and a decision variable definition unit;

[0094] The bridge basic information unit is used to store and manage the bridge's design parameters, including structural type, material properties, geometric dimensions, design loads, and environmental conditions. The cost calculation unit is used to calculate various costs, including initial construction costs, periodic inspection costs, maintenance costs at all levels, failure loss costs, and corresponding discounting methods. The degradation modeling unit is used to describe the stochastic degradation process of the bridge's technical condition level. The maintenance effect modeling unit is used to define the improvement results and duration of different maintenance measures on the bridge's technical condition level. The life cycle time management unit is used to set and manage the bridge's service life time range and provide discretization processing for time nodes. The decision variable definition unit is used to define the decision variables in the genetic algorithm, including inspection time, maintenance time, and maintenance method selection.

[0095] Furthermore, the process of acquiring the optimization process data includes:

[0096] The parameters of the preset bridge life cycle model are initialized, including population size, crossover rate, mutation rate, and maximum number of iterations. An initial population is generated using the bridge basic information unit and the decision variable definition unit, where each individual in the initial population represents a complete life cycle maintenance strategy. Monte Carlo simulation is performed on each individual in the initial population, using the degradation modeling unit to simulate the bridge state evolution process and the cost calculation unit to calculate the total life cycle cost. During the genetic algorithm iteration process, intermediate data at each fitness evaluation are recorded, including time t, bridge state S, and the expected life cycle cost from time t to the end of the life cycle. After multiple generations of evolution, a large amount of sample data covering different times and states is collected to form optimization process data.

[0097] Furthermore, the terminal cost value function generation module is used to generate a terminal cost value function using the optimization process data to estimate the expected cost of the bridge's remaining life when the bridge is in state S at a future time t.

[0098] The process of generating the terminal cost value function includes:

[0099] Sample data is extracted from the optimization process data; the sample data is fitted using machine learning methods; an initial terminal cost value function is generated based on the fitted sample data, and the initial terminal cost value function is verified and calibrated to obtain the terminal cost value function. The expected cost of the remaining lifetime period at time t and state S is calculated to obtain the expected cost of the remaining lifetime.

[0100] The process of generating the terminal cost value function further includes a deep learning optimization step:

[0101] A deep neural network model is constructed, including an input layer, multiple hidden layers, and an output layer. The input layer receives combined features of time t and bridge state S. The hidden layers employ ReLU activation function and Dropout regularization. The output layer predicts the expected cost of remaining lifetime. The network is trained using Adam optimizer and mean squared error loss function. Training stability is improved through batch normalization and learning rate decay strategies.

[0102] An ensemble learning mechanism was established to weightedly fuse the prediction results of support vector regression, random forest, and deep neural networks, with the weights dynamically adjusted based on validation set performance. By introducing deep learning technology to optimize the terminal cost value function, and employing advanced methods such as deep neural network models and the ensemble learning mechanism, significant improvements in the prediction accuracy and generalization ability of the value function were achieved. This ensemble learning mechanism fully leverages the powerful capabilities of deep learning in modeling complex nonlinear relationships. A multi-layered neural network structure captures the deep-seated correlation patterns between bridge state, time factors, and costs. The application of ReLU activation function and Dropout regularization effectively avoids overfitting, while the Adam optimizer and learning rate decay strategy ensure the stability and convergence of model training. The establishment of the ensemble learning mechanism further enhances the system's robustness. By intelligently fusing the prediction results of multiple methods such as support vector regression, random forest, and deep neural networks, not only is prediction accuracy improved, but the adaptability to abnormal situations is also enhanced, providing a more solid technical guarantee for the reliable operation of the dynamic optimization system in complex engineering environments.

[0103] Furthermore, the dynamic rolling optimization module is used to perform rolling optimization at decision points during the bridge's lifespan, and includes: a rolling optimization window determination unit and a rolling optimization problem construction unit;

[0104] The rolling optimization window determination unit acquires the actual bridge state data at the decision time, assesses the bridge risk based on the actual bridge state data, and determines the rolling optimization window; the determination process of the rolling optimization window includes:

[0105] Acquire the actual bridge status data at the decision-making time, including the time information of the current decision-making time, the bridge technical condition level, the actual technical condition parameters of each bridge component, recent inspection history records, historical degradation trend information, and changes in environmental factors and load conditions; conduct a risk assessment of the bridge based on the actual bridge status data to obtain the bridge risk level.

[0106] If the bridge risk level is in the low-risk range, the optimized window length is set to [N, N+T] years; if the bridge risk level is in the medium-risk range, the optimized window length is set to [NM, N) years; if the bridge risk level is in the high-risk range, the optimized window length is set to [NP, NM) years; where 0 <N-P<N-M<N<N+T;

[0107] The time interval of the scrolling optimization window is determined based on the length of each optimized window.

[0108] The process of determining the scrolling optimization window further includes a multi-level risk assessment mechanism:

[0109] Establish a structural layer risk assessment, and calculate the component risk index based on the technical status parameters of major components such as main beams, piers, and foundations; establish a system layer risk assessment, and calculate the system risk index by comprehensively considering the overall structural performance, traffic function performance, and safety performance of the bridge; establish a network layer risk assessment, and calculate the network risk index by considering the importance of the bridge in the road network, the availability of alternative routes, and socio-economic impact.

[0110] The Analytic Hierarchy Process (AHP) is used to determine the risk weights of multiple levels, and a comprehensive risk assessment model is constructed. The window length is dynamically adjusted according to the comprehensive risk level to achieve refined window management.

[0111] By establishing a multi-level risk assessment mechanism encompassing structural, system, and network layers, and employing the analytic hierarchy process (AHP) for weight determination, the comprehensiveness, systematicity, and scientific rigor of bridge risk assessment have been significantly enhanced. This multi-level risk assessment mechanism expands traditional single-dimensional risk assessment into a comprehensive multi-level assessment system. Structural layer risk assessment focuses on the technical condition of the bridge's physical components; system layer risk assessment comprehensively considers the overall performance of the bridge; and network layer risk assessment focuses on the bridge's importance and social impact within the transportation network. These three levels complement and verify each other, forming a three-dimensional risk identification and quantification system. The application of the AHP ensures the scientific and rational allocation of multi-level risk weights, while the refined dynamic window adjustment mechanism enables the system to respond more accurately to subtle changes in the bridge's risk status. This refined risk management not only improves the targeting and effectiveness of maintenance decisions but also provides crucial technical support for the intelligent and precise management of bridge assets.

[0112] Furthermore, the rolling optimization problem construction unit constructs a short-term optimization problem within the rolling optimization window by combining the expected cost of remaining lifetime, and solves the short-term optimization problem through a genetic algorithm to obtain the optimal action sequence within the rolling optimization window;

[0113] The short-run optimization problem, which incorporates the expected cost of remaining lifetime into a rolling optimization window, is as follows:

[0114] The objective function of the short-run optimization problem is set as the sum of the in-window cost and the terminal cost; wherein, the in-window cost is the simulated expected cost generated by the maintenance strategy simulation within the rolling optimization window, and the terminal cost is the expected cost of the bridge's remaining life at the end of the window, calculated by calling the terminal cost value function.

[0115] Define the decision variables for the short-term optimization problem; the decision variables are represented using variable precision.

[0116] The process of constructing the short-run optimization problem further includes a robust optimization mechanism:

[0117] Establish a set of parameter uncertainties, including uncertainties in degradation rate, maintenance costs, and discount rate;

[0118] A box-type uncertainty set is used to describe the range of parameter variation, and the uncertainty boundary is determined by historical data statistics and expert judgment. A robust optimization objective function is constructed, and an adversarial genetic algorithm is used to solve the robust optimization problem. Max-min bilevel optimization is used to ensure that the decision still has good performance in the worst case.

[0119] By establishing a robust optimization mechanism that considers parameter uncertainties, and employing a box-type uncertainty set description and an adversarial genetic algorithm for solution, the reliability and robustness of dynamic optimization systems under uncertain environments are significantly enhanced. This robust optimization mechanism fully considers the parameter uncertainties widely present in practical engineering, including fluctuations in degradation rate, changes in maintenance costs, and adjustments to the discount rate. By constructing an uncertainty set and a max-min bilevel optimization model, it ensures that the optimized maintenance strategy still exhibits good performance even in the worst-case scenario. The application of the adversarial genetic algorithm enables efficient solution to the robust optimization problem. By simulating adversarial changes in uncertain parameters, it seeks maintenance strategies that perform well in all possible situations. This robust design not only improves the reliability and safety of decision-making schemes but also enhances the system's adaptability to complex and changing engineering environments, providing important technical support for scientific decision-making in bridge maintenance management under uncertain conditions.

[0120] Furthermore, the process of obtaining the optimal action sequence includes:

[0121] A genetic algorithm is run on the constructed short-term optimization problem to iteratively optimize through selection, crossover, and mutation operations, generating an initial population representing different action sequences within a rolling optimization window. Degenerate modeling units and Monte Carlo simulations are used to calculate the in-window cost of each individual in the initial population within the rolling optimization window. The terminal cost value function is called to calculate the terminal cost of each individual. The objective function of the short-term optimization problem is used as the fitness function to evaluate the quality of each individual in the population and select the optimal individual. When the convergence condition is met, the action sequence corresponding to the optimal individual is output, obtaining the optimal action sequence. The optimal action sequence contains a complete maintenance and inspection plan within the rolling optimization window, including inspection time, maintenance method, and maintenance time.

[0122] Furthermore, the rolling decision output module is used to extract only the first action instruction from the optimal action sequence and output it as the current optimal maintenance decision.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm, characterized in that, include: The initialization module is used to perform a genetic algorithm optimization covering the entire life cycle of a bridge at the initial stage of its life cycle, based on a preset full life cycle model of the bridge, to obtain optimization process data. The terminal cost value function generation module is used to generate a terminal cost value function using the optimization process data to estimate the expected cost of the bridge's remaining life when the bridge is in state S at future time t. The dynamic rolling optimization module is used to perform rolling optimization at decision points throughout the bridge's lifespan, including: The rolling optimization window determination unit acquires the actual bridge status data at the decision time, assesses the bridge risk based on the actual bridge status data, and determines the rolling optimization window. The rolling optimization problem construction unit combines the remaining lifetime expected cost to construct a short-term optimization problem within the rolling optimization window, and solves the short-term optimization problem through a genetic algorithm to obtain the optimal action sequence within the rolling optimization window; The rolling decision output module is used to extract only the first action instruction from the optimal action sequence and output it as the current optimal maintenance decision.

2. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 1, characterized in that: The preset bridge life cycle model includes: a bridge basic information unit, a cost calculation unit, a degradation modeling unit, a maintenance effect model unit, a life cycle time management unit, and a decision variable definition unit; The bridge basic information unit is used to store and manage the bridge's design parameters, including structural type, material properties, geometric dimensions, design loads, and environmental conditions. The cost calculation unit is used to calculate various costs, including initial construction costs, periodic inspection costs, maintenance costs at all levels, failure loss costs, and corresponding discounting methods. The degradation modeling unit is used to describe the stochastic degradation process of the bridge's technical condition level. The maintenance effect modeling unit is used to define the improvement results and duration of different maintenance measures on the bridge's technical condition level. The life cycle time management unit is used to set and manage the bridge's service life time range and provide discretization processing for time nodes. The decision variable definition unit is used to define the decision variables in the genetic algorithm, including inspection time, maintenance time, and maintenance method selection.

3. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 2, characterized in that: The process of acquiring optimization process data includes: The parameters of the preset bridge life cycle model are initialized, including population size, crossover rate, mutation rate, and maximum number of iterations. An initial population is generated using the bridge basic information unit and the decision variable definition unit, where each individual in the initial population represents a complete life cycle maintenance strategy. Monte Carlo simulation is performed on each individual in the initial population, using the degradation modeling unit to simulate the bridge state evolution process and the cost calculation unit to calculate the total life cycle cost. During the genetic algorithm iteration process, intermediate data at each fitness evaluation is recorded, including the bridge state S at time t and the expected life cycle cost from time t to the end of the life cycle. After multiple generations of evolution, the intermediate data covering different times and states are collected to form optimization process data.

4. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 3, characterized in that: The process of generating the terminal cost value function includes: Sample data is extracted from the optimization process data; the sample data is fitted using machine learning methods; an initial terminal cost value function is generated based on the fitted sample data, and the initial terminal cost value function is verified and calibrated to obtain the terminal cost value function. The expected cost of the remaining lifetime period at time t and state S is calculated to obtain the expected cost of the remaining lifetime.

5. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 1, characterized in that: The process of determining the scrolling optimized window includes: Acquire the actual bridge status data at the decision-making time, including the time information of the current decision-making time, the bridge technical condition level, the actual technical condition parameters of each bridge component, recent inspection history records, historical degradation trend information, and changes in environmental factors and load conditions; conduct a risk assessment of the bridge based on the actual bridge status data to obtain the bridge risk level. If the bridge risk level is in the low-risk range, the optimized window length is set to [N, N+T] years; if the bridge risk level is in the medium-risk range, the optimized window length is set to [NM, N) years; if the bridge risk level is in the high-risk range, the optimized window length is set to [NP, NM) years; where 0 <N-P<N-M<N<N+T; The time interval of the scrolling optimization window is determined based on the length of each optimized window.

6. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 1, characterized in that: The short-run optimization problem, which incorporates the expected cost of remaining lifetime into a rolling optimization window, is as follows: The objective function of the short-run optimization problem is set as the sum of the in-window cost and the terminal cost; wherein the in-window cost is the simulated expected cost generated by the maintenance strategy simulation within the rolling optimization window, and the terminal cost is the expected cost of the bridge's remaining life at the end of the window calculated by calling the terminal cost value function. Define the decision variables for the short-term optimization problem; the decision variables are represented using variable precision.

7. The dynamic optimization system for bridge life-cycle maintenance strategy based on genetic algorithm according to claim 1, characterized in that: The process of obtaining the optimal action sequence includes: A genetic algorithm is run on the constructed short-term optimization problem to iteratively optimize through selection, crossover, and mutation operations, generating an initial population representing different action sequences within a rolling optimization window. Degenerate modeling units and Monte Carlo simulations are used to calculate the in-window cost of each individual in the initial population within the rolling optimization window. The terminal cost value function is called to calculate the terminal cost of each individual. The objective function of the short-term optimization problem is used as the fitness function to evaluate the quality of each individual in the population and select the optimal individual. When the convergence condition is met, the action sequence corresponding to the optimal individual is output, obtaining the optimal action sequence. The optimal action sequence contains a complete maintenance and inspection plan within the rolling optimization window, including inspection time, maintenance method, and maintenance time.

Images