Self-adaptive optimization decision-making method for preventive maintenance opportunity of road surface

Abstract

The invention relates to a self-adaptive optimization decision-making method for preventive maintenance opportunity of a road surface, which comprises the following steps of: establishing a road surface performance prediction model based on multi-source data fusion based on historical road surface performance data, traffic load, climate environment and material structure characteristics; based on the performance data of the pavement before and after maintenance construction, extracting the instantaneous performance resilience value after maintenance completion and the performance attenuation rate after maintenance, and establishing a maintenance effect prediction model; constructing a double-layer optimization decision model taking the total cost minimization of the whole life cycle as an optimization target; the upper layer takes the maintenance opportunity threshold value as a decision variable to generate a corresponding maintenance demand; the lower layer solves the optimal maintenance schedule under the corresponding threshold value; in the lower-layer solving process, a maintenance effect prediction model is called to predict the maintenance effect; and an optimal maintenance opportunity threshold value is obtained through simulation optimization search. The method aims to realize closed-loop coupling of performance prediction, maintenance effect quantification and maintenance decision, and minimize the total cost of the whole life cycle on the premise of ensuring the pavement service level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pavement maintenance decision optimization, and particularly relates to a self-adaptive optimization decision method for pavement preventive maintenance timing. BACKGROUND

[0002] As a core component of highway infrastructure, pavement directly bears traffic load and determines service level, and its performance degradation rate directly affects road operation safety, driving comfort and life cycle cost. Preventive maintenance, as a key strategy of modern highway maintenance management, aims to implement low-cost maintenance measures before significant damage occurs in the pavement structure, so as to delay performance degradation, extend service life and reduce life cycle cost. Scientific determination of preventive maintenance timing is the core to achieve the goal of “life extension and cost reduction”, which can not only avoid waste of resources caused by early intervention, but also prevent structural damage and functional failure caused by late treatment, and has significant engineering and economic value.

[0003] In recent years, significant progress has been made in pavement performance prediction research. Research methods have developed from traditional empirical models (such as the HDM model of the World Bank and the LTPP performance model of the United States) to multi-source data prediction models based on machine learning, significantly improving prediction accuracy and applicability. In terms of performance threshold and maintenance effect modeling, research has shifted from static threshold based on a single indicator (such as the International Roughness Index IRI) to dynamic models based on multi-index coupling; some research has preliminarily realized the dynamic characterization of maintenance effect by analyzing the performance recovery law before and after maintenance. In terms of maintenance optimization decision-making, methods have evolved from empirical solutions based on life cycle cost analysis (LCCA) to multi-objective decision-making models using intelligent algorithms such as genetic algorithm and particle swarm optimization, which can provide near-optimal solutions under budget and resource constraints. However, existing research still has the following limitations: first, most methods stop at “prediction” or “evaluation”, and fail to achieve closed-loop coupling of performance prediction and threshold optimization; second, the quantitative characterization of maintenance effect is insufficient, making it difficult to accurately reflect the recovery differences under different measures and road conditions; third, cost models are mostly based on foreign parameter systems, lacking localization calibration and constraint integration for China's climate and traffic characteristics. These shortcomings collectively limit the adaptive ability and engineering applicability of maintenance timing decision-making. Therefore, there is an urgent need for a new decision-making method that can integrate multi-source data, finely quantify maintenance effect, and automatically and efficiently output the optimal maintenance timing under actual constraints. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a self-adaptive optimization decision method for pavement preventive maintenance timing, which aims to achieve closed-loop coupling of performance prediction, maintenance effect quantification and maintenance decision-making, and adaptively search for the optimal maintenance timing threshold, thereby minimizing the total life cycle cost while ensuring the service level of the pavement.

[0005] To achieve the above object, the application adopts the following technical scheme: a self-adaptive optimization decision method for road surface preventive maintenance timing, comprising:

[0006] S1, based on historical road surface performance data, traffic load characteristics, climate environment characteristics and material structure characteristics, a multi-source data fusion road surface performance prediction model is established by using machine learning algorithm, so as to predict the road surface performance at future time points by using the road surface performance prediction model;

[0007] S2, based on the performance data of road surface before and after maintenance construction, the parameters representing the maintenance effect are extracted, including the instantaneous performance rebound amount after maintenance and the performance decay rate after maintenance, and a maintenance effect prediction model is established, so as to predict the maintenance effect by using the maintenance effect prediction model;

[0008] S3, a double-layer optimization decision model is constructed with the minimum total cost of the whole life cycle as the optimization objective; the upper layer takes the maintenance timing threshold as the decision variable, and when the road surface performance predicted by the road surface performance prediction model deteriorates to the threshold , the corresponding maintenance demand is generated; the lower layer solves the optimal maintenance scheduling of the maintenance demand under the corresponding threshold ; in the lower layer solving process, the maintenance effect prediction model is called to predict the maintenance effect, and then the future deterioration path of the road surface performance is updated; then the optimization algorithm is used for simulation optimization, and the optimal maintenance timing threshold is adaptively searched.

[0009] Further, the implementation method of the road surface performance prediction model is:

[0010] Step A1: input historical performance data set and XGBoost hyperparameters; the historical performance data set includes road surface performance data, time-varying characteristics and static characteristics of the i-th road section at different time points, the time-varying characteristics are composed of traffic load characteristics and climate environment characteristics, and the static characteristics are material structure characteristics;

[0011] Step A2: adopt sliding time window to construct training samples, each training sample contains road surface performance data at the previous time point of the i-th road section , time-varying characteristics at the current time point of the i-th road section , static characteristics at the current time point of the i-th road section and road surface performance data at the current time point of the i-th road section ;

[0012] Step A3: Perform data preprocessing on the training samples;

[0013] Step A4: Based on the XGBoost algorithm, train and optimize the model parameters using training samples to minimize the following objective function. :

[0014]

[0015]

[0016] in, The number of training samples. For pavement performance prediction results, is the pavement performance prediction function, which is a regression model trained by the XGBoost algorithm, used to find the mapping relationship between input features and output data;

[0017] Step A5: Output the pavement performance prediction function This allows us to use time-varying and static characteristics of future time points, as well as road performance data prior to those future time points, to predict road performance at future time points.

[0018] Furthermore, the current time point of the i-th road segment pavement performance data , , , Representing the current time point of the i-th segment respectively The smoothness index, rut depth index, and crack damage index.

[0019] Furthermore, in step A3, data preprocessing is performed on the training samples, specifically including:

[0020] Outlier handling: Use Z-score to remove outliers in continuous variables, and use box plots to remove outliers in categorical or proportional variables;

[0021] Missing value imputation: Linear interpolation is used to impute missing data in time series features, the KNN algorithm is used to impute sequences with large fluctuations, and the mode within a group is used to impute missing structural features;

[0022] Standardization: All features are uniformly resampled to a quarterly resolution, and Min-Max Scaling is used to standardize the data to the range [0, 1].

[0023] After training in step A4, the SHAP analysis framework is used to perform interpretability analysis on the pavement performance prediction model, calculate the average Shapley value of each feature, and quantify its contribution to pavement performance prediction results. The global contribution is calculated, and feature importance ranking and dependency graphs at the road segment and network levels are generated.

[0024] Furthermore, suppose the first This section of road in the quarter Maintenance work was carried out, and If maintenance work is completed and pavement performance testing is conducted at the end of the quarter, the pavement performance evolution model is as follows:

[0025]

[0026] in, The pavement performance values ​​obtained from the first test after maintenance are completed; To achieve instantaneous performance rebound during maintenance, among which This indicates the quarter of the last inspection before construction. This indicates the first quarterly inspection after the completion of maintenance and construction. The rate of performance degradation after maintenance;

[0027] Thus, the pavement performance time series of the i-th road segment is obtained. .

[0028] Furthermore, the implementation method of the maintenance effect prediction model is as follows:

[0029] Step B1: Input pavement performance time series The first quarterly inspection after maintenance and construction is completed. Types of measures Time-varying characteristics and static features ;

[0030] Step B2: Construct the analysis window ;

[0031] Step B3: Perform data preprocessing on the input data;

[0032] Step B4: Within the post-curing window, calculate the instantaneous performance rebound after curing is completed. and the rate of performance degradation after maintenance ;

[0033] Step B5: Construct the feature dataset, where the feature data group is... ;

[0034] Step B6: Input feature data { },by To achieve the regression objective, the maintenance effect prediction model is trained to minimize the weighted mean square error. :

[0035]

[0036]

[0037] in, The amount of feature data used for training, The coefficient is used to balance the differences in dimensions. A function for predicting maintenance effects;

[0038] Step B7: Output the maintenance effect prediction function To predict the maintenance effect.

[0039] Furthermore, in step B3, the input data is preprocessed, including applying Hampel filtering to identify and remove outliers to achieve data cleaning, and linearly interpolating missing values.

[0040] In step B4, the instantaneous performance rebound after curing is calculated using the following formula. :

[0041]

[0042] in, Represents the median operator;

[0043] In the post-maintenance area Internally, robust linear regression fitting is used. :

[0044] .

[0045] Furthermore, the implementation method of the two-level optimization decision model is as follows:

[0046] Step C1: Input the pavement performance prediction function Maintenance effect prediction function Cost parameters and constraints;

[0047] Step C2: Initialize the design: in the policy space The initial sample set for generating maintenance timing thresholds is generated using Latin hypercube sampling. ,in The number of samples;

[0048] Step C3: For each sample Perform steps C4-C7;

[0049] Step C4: Based on samples and pavement performance prediction function Conduct Monte Carlo simulations to generate maintenance requirements for the planning period, i.e., through... Natural degradation simulation was performed on the pavement performance of this road section. When the pavement performance predicted by the pavement performance prediction model deteriorated to a threshold... When this happens, maintenance is required;

[0050] Step C5: After the maintenance requirement is triggered, with the goal of minimizing the total cost over the entire life cycle, a genetic algorithm is used to solve the scheduling subproblem under the threshold θ, optimizing the implementation sequence and combination of measures for the maintenance requirement to obtain the optimal maintenance schedule. During the solution process, the maintenance effect prediction function is invoked each time a maintenance measure is executed. Predict the effects of maintenance and then update the future deterioration path of pavement performance;

[0051] Step C6: Calculate the net present value of the total life-cycle cost:

[0052]

[0053] in, It represents the net present value of the total life-cycle cost; Represents the net present value cost function; Indicates at the threshold and maintenance schedule The institutional costs below; Indicates at the threshold and maintenance schedule Lower user costs;

[0054] Step C7: If Greater than the set maximum allowable limit Then directly to Imposing extremely high penalties , that is to say This enables the proxy model to identify the region as an infeasible region;

[0055] Step C8: Based on the dataset Training the Gaussian process surrogate model: ;

[0056] Step C9: Continue performing the following operations until the number of iterations reaches the upper limit or the improvement rate falls below the set threshold. Using the expected improvement criterion, in Select the next sampling point and to Perform steps C3-C7 to expand the sample set and refit. Model;

[0057] Step C10: Output the optimal maintenance timing threshold .

[0058] Furthermore, in step C5, the objective function for solving the lower-level scheduling optimization problem is expressed as:

[0059]

[0060] in, This indicates the optimal maintenance schedule; Indicates at the threshold and maintenance schedule The institutional costs below; Indicates at the threshold and maintenance schedule Lower user costs; This represents the feasible region.

[0061] Furthermore, feasible domain Maintenance timing threshold The annual maintenance requirements and constraints that trigger the process are determined; these constraints include:

[0062] 1) Performance constraints, the expression of which is:

[0063]

[0064] 2) Annual budget constraint, its expression is:

[0065]

[0066] in, This is the annual budget cap;

[0067] 3) Seasonal construction constraints, expressed as:

[0068]

[0069] That is, for all periods of winter. For maintenance operations such as "micro-surfacing" and "fog sealing", the maintenance schedule must be 0, meaning that construction is not allowed.

[0070] 4) Construction period and traffic organization constraints, expressed as:

[0071]

[0072] in, The maximum available construction period for the quarter. This represents the upper limit on the number of projects that can be run in parallel.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] 1) Closed-loop decision-making and adaptive optimization: This invention is the first to place performance prediction, maintenance effect quantification, and threshold optimization within a unified closed-loop framework. Through a two-layer optimization model, the maintenance strategy is automatically and dynamically adjusted based on road condition prediction and cost targets, overcoming the problem of disconnect between prediction and decision-making in traditional methods.

[0075] 2) Refined and quantified maintenance effect: The proposed “Δ + k’” model refines the maintenance effect from two dimensions: “instantaneous rebound” and “long-term decay”. Compared with the traditional fixed recovery coefficient, it can more realistically reflect the comprehensive performance of different measures under different conditions, and provide more accurate input for optimization decision-making.

[0076] 3) High decision-making efficiency: For complex two-layer black-box optimization problems, an optimization algorithm based on a sequence surrogate model is designed. By replacing expensive full-process simulation with intelligent sampling and surrogate model, the computational cost is greatly reduced, making the method applicable to optimization decision-making of large-scale road networks. Attached Figure Description

[0077] Figure 1 This is a flowchart illustrating the implementation of the adaptive optimization decision-making method for the timing of preventive road maintenance provided in this embodiment of the invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0079] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0080] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0081] like Figure 1 As shown, this embodiment provides an adaptive optimization decision-making method for the timing of preventive road maintenance, including the following steps:

[0082] S1. Based on historical pavement performance data, traffic load characteristics, climate environment characteristics and material structure characteristics, a pavement performance prediction model is established by using machine learning algorithms to fuse multi-source data, so as to predict the pavement performance at future time points using the pavement performance prediction model.

[0083] S2. Based on the performance data of the road surface before and after maintenance construction, extract parameters that characterize the maintenance effect, including the instantaneous performance rebound after maintenance and the performance decay rate after maintenance, and establish a maintenance effect prediction model to predict the maintenance effect.

[0084] S3. Construct a two-level optimization decision model with the goal of minimizing the total cost over the entire life cycle; the upper level uses maintenance timing thresholds. As a decision variable, when the pavement performance predicted by the pavement performance prediction model deteriorates to a threshold... At that time, the corresponding maintenance requirements are generated; the lower layer solves for the corresponding thresholds. The optimal maintenance schedule for this maintenance requirement. In the lower-level solution process, a maintenance effect prediction model is invoked to predict the maintenance effect, thereby updating the future deterioration path of pavement performance; then, an optimization algorithm is used for simulation optimization, and the optimal maintenance timing threshold is obtained through adaptive search. .

[0085] 1 Performance Prediction Model

[0086] 1.1 Modeling Approach

[0087] The time-varying degradation process of pavement performance is jointly influenced by traffic load, climate environment, and material properties, exhibiting significant nonlinearity and regional variability. While traditional empirical equation models (such as HDM-4) possess strong interpretability, they struggle to characterize the high-order coupling relationships among multiple factors, leading to insufficient prediction accuracy. Therefore, this paper proposes a machine learning performance prediction model that integrates traffic, climate, and material characteristics. Using a quarterly time step, it establishes dynamic regression relationships among multiple indicators to achieve long-term accurate prediction of pavement performance. The model concept is as follows:

[0088]

[0089] in, For road section In the quarter The performance state vector, , , Representing the current time point of the i-th segment respectively Smoothness index, rut depth index, and crack damage index; Features that vary over time (such as traffic and climate); These are static characteristics (material type, thickness, base layer form, etc.). This is a nonlinear mapping function based on ensemble learning.

[0090] 1.2 Input Features and Data Preprocessing

[0091] This method integrates three types of multi-source data features as inputs to the pavement performance prediction model: traffic load features, climate environment features, and material structure features. The specific variables, data sources, and theoretical basis for each feature are shown in Table 1.

[0092] Traffic load characteristics ( This data is used to quantify the traffic stress experienced by a road during operation. Key variables include average daily traffic volume (AADT), the proportion of loaded vehicles, and equivalent standard axle loads (ESALs). ESALs are the cumulative value of traffic volume from different axle types and loads converted to standard axle loads, and are a direct driver of pavement mechanical response. This data primarily comes from long-term monitoring records from the Dynamic Weighing Information System (WIM) and the Road Management System (PMS) database.

[0093] Climate and environmental characteristics ( (This refers to the time-varying influence of external climatic conditions on the properties of pavement materials, such as the stiffness modulus of asphalt mixtures and the moisture content of the base soil. It includes seasonal average temperature, annual temperature range, cumulative precipitation, and the number of freeze-thaw cycles. These data are derived from observational data from national and provincial meteorological stations.)

[0094] Material structural characteristics ( ): As a static attribute variable, it is used to describe the inherent physical and mechanical properties of the pavement structure itself, including surface layer type, surface layer thickness, base layer type, and modulus parameters. This information is extracted from the project's early engineering ledger and design documents.

[0095] Table 1 Input characteristics of the pavement performance prediction model

[0096] Characteristic type Example variable Data source Description Traffic load characteristics ) Annual average daily traffic, heavy vehicle percentage, equivalent standard axle load Dynamic weighing, pavement management system database Characterizes external load actions, main cause of fatigue damage and rutting. Climate environment characteristics ) Seasonal average temperature, temperature annual range, precipitation, freeze-thaw cycles National / provincial weather stations Environmental driving factors that affect material physical-mechanical properties (e.g. modulus, moisture). Material structure features ) Surface course type, thickness, base course type, modulus parameters Engineering records, design documents Characterizes internal structure resistance, determines its response to load and environmental actions

[0097] During the data preprocessing stage, targeted cleaning and standardization strategies were adopted for different data types and characteristics to ensure data quality and model stability.

[0098] Outlier handling: For continuous variables in traffic load and climate environment (such as ESALs and temperature), the Z-score criterion (with a threshold set at ±3) is used to identify and remove observations that deviate significantly. For categorical or proportional variables (such as the proportion of heavy vehicles), the box plot method is used to identify and remove data points that exceed 1.5 times the upper or lower quartile range as outliers.

[0099] Missing value imputation: For missing data in time series variables (such as ESALs, precipitation), linear interpolation is preferred for imputation; for nonlinear or highly fluctuating series, the K-nearest neighbor (KNN) algorithm is used for imputation to maintain the integrity of the time series pattern. For missing static features (such as base course category, surface layer type), they are grouped according to their respective engineering sections, and the mode of that group is used for imputation.

[0100] Data alignment and standardization: To ensure consistency across all time-series features, all variables are uniformly resampled to a quarterly resolution. Finally, to eliminate the negative impact of different variable scales on model training, all input features are normalized to the range [0, 1] (Min-Max Scaling). The complete, consistent, and standardized feature matrix obtained after the above process is used as the final input to the model.

[0101] 1.3 Algorithm Implementation of Road Performance Prediction Model

[0102] The pavement performance prediction model takes historical performance data, multi-source features, and pavement structural attributes as input, constructs samples through a time-series sliding window, and uses the XGBoost algorithm for training and prediction. The specific implementation process of the pavement performance prediction model is as follows.

[0103] Step A1: Input historical performance dataset The XGBoost hyperparameters include: the historical performance dataset, which includes pavement performance data, time-varying features, and static features at different time points for the i-th road segment; the time-varying features consist of traffic load features and climate environment features; and the static features are material structure features. The XGBoost hyperparameters include: maximum tree depth. Learning rate η, number of trees and sliding time window width .

[0104] Step A2: Construct training samples using a sliding time window. Each training sample consists of data from two adjacent time points, and each training sample includes the time point preceding the i-th road segment. pavement performance data The current time point of the i-th segment Time-varying characteristics The current time point of the i-th segment static features and the current time point of segment i pavement performance data .

[0105] Step A3: Perform data preprocessing on the training samples, specifically including:

[0106] Outlier handling: Z-score (with a threshold of ±3) is used to remove outliers in continuous variables, and box plots are used to remove outliers in categorical or proportional variables.

[0107] Missing value imputation: Linear interpolation is used to impute missing data in time series features, the KNN algorithm is used to impute sequences with large fluctuations, and the mode within a group is used to impute missing structural features;

[0108] Standardization: All features are uniformly resampled to quarterly resolution, and Min-Max Scaling is used to standardize the data to the range [0, 1].

[0109] Step A4: Based on the XGBoost algorithm, train and optimize the model parameters using training samples to minimize the following objective function. :

[0110]

[0111]

[0112] in, The number of training samples. For pavement performance prediction results, The pavement performance prediction function is a regression model trained by the XGBoost algorithm, used to find the mapping relationship between input features and output data.

[0113] Step A5: Use the SHAP analysis framework to perform interpretability analysis on the pavement performance prediction model, calculate the average Shapley value of each feature, and quantify its contribution to pavement performance prediction results. The global contribution is calculated, and feature importance ranking and dependency graphs at the road segment and network levels are generated.

[0114] Step 6: Output the road surface performance prediction function This allows us to use time-varying and static characteristics of future time points, as well as road performance data prior to those future time points, to predict road performance at future time points.

[0115] Through the algorithms described above, the pavement performance prediction model can quantify the impact of traffic, climate, and material factors on pavement degradation, providing a precise basis for subsequent preventative maintenance decisions. Furthermore, the model's interpretability analysis helps decision-makers better understand the influence of different factors, thereby enabling them to make more scientific and rational maintenance decisions.

[0116] 2. Quantitative Model of Maintenance Effect

[0117] 2.1 Model Idea

[0118] The implementation of maintenance measures will interrupt the natural decline trajectory of pavement performance, forming a typical pattern of "performance jump – re-decay". In order to achieve closed-loop optimization from performance prediction to maintenance decision-making, it is necessary to accurately characterize the quantitative effect of maintenance intervention. This paper proposes a "Δ + k′" parameterized model to decouple the maintenance effect into two core dimensions: (1) Instantaneous recovery benefit Δ: reflects the immediate rebound amplitude of pavement performance after maintenance construction. (2) Long-term durability performance k′: characterizes the performance decline rate after maintenance and measures the long-term maintenance capability of the measures.

[0119] Assume the first This section of road in the quarter Maintenance work was carried out, and If maintenance work is completed and pavement performance testing is conducted at the end of the quarter, the pavement performance evolution model is as follows:

[0120]

[0121] in, The pavement performance values ​​obtained from the first test after maintenance are completed; To achieve instantaneous performance rebound during maintenance, among which This indicates the quarter of the last inspection before construction. This indicates the first quarterly inspection after the completion of maintenance and construction. The model represents the rate of performance degradation after maintenance. It naturally aligns with measured data and comprehensively characterizes the overall effect of maintenance measures from two perspectives: "amplitude of change" and "rate of degradation".

[0122] Thus, the pavement performance time series of the i-th road segment is obtained. .

[0123] 2.2 Parameter Extraction Method

[0124] Based on road segment Quarterly performance timings The parameter extraction process is as follows:

[0125] (1) Track window selection

[0126] Around the time of maintenance completion Select the one containing the previous Quarter and after Quarterly analysis window: .

[0127] (2) Instantaneous rebound Extraction

[0128] To robustly estimate performance jumps, we define:

[0129]

[0130] in, The median operator can smooth out occasional fluctuations; for samples with abnormal climates (heavy rain, high temperature, etc.), Hampel filter or Huber regression can be used to remove them.

[0131] (3) New decay rate Fitting

[0132] In the post-maintenance area Internally, a robust linear regression fit is used:

[0133]

[0134] If the Bayesian Information Criterion (BIC) test confirms the existence of significant nonlinearity, then an ARIMA or Exponential Smoothing (ETS) model is used for fitting, and the long-term trend slope is taken as the... The estimate.

[0135] (4) Data quality control

[0136] Road segment data that are missing for ≥3 consecutive quarters or have an outlier rate >20% are used only for extrapolation validation and are not used for training.

[0137] 2.3 Association Modeling

[0138] Extracted parameters The digital twin of maintenance effectiveness is closely related to the type of maintenance measures, structural properties, and environmental conditions. Establish a maintenance effectiveness prediction model:

[0139]

[0140] in, Types of maintenance measures (micro-surfacing, ultra-thin wear layer, milling overlay, etc.); Multiple linear regression, BP neural network, or LSTM network can be used; the model output can generalize to predict the maintenance effect under any given situation, providing core input for threshold optimization.

[0141] 2.4 Algorithm Implementation of the Maintenance Effect Prediction Model

[0142] To achieve a systematic process for quantifying and predicting maintenance effects, an algorithm for extracting and modeling maintenance effects is proposed to realize the maintenance effect prediction model. The specific implementation process of the maintenance effect prediction model is as follows.

[0143] Step B1: Input pavement performance time series The first quarterly inspection after maintenance and construction is completed. Types of measures Time-varying characteristics and static features .

[0144] Step B2: Window Delineation: Constructing the Analysis Window .

[0145] Step B3: Perform data preprocessing on the input data, including: applying Hampel filtering to identify and remove outliers to achieve data cleaning, and linearly imputing missing values.

[0146] Step B4: Within the post-curing window, calculate the instantaneous performance rebound after curing is completed. and the rate of performance degradation after maintenance .

[0147] The instantaneous performance rebound after maintenance is calculated using the following formula. :

[0148]

[0149] in, This represents the median operator.

[0150] In the post-maintenance area Internally, robust linear regression fitting is used. :

[0151] .

[0152] Step B5: Construct the feature dataset and record the feature data groups. .

[0153] Step B6: Input feature data { },by To achieve the regression objective, the maintenance effect prediction model is trained to minimize the weighted mean square error. :

[0154]

[0155]

[0156] in, The amount of feature data used for training, The coefficient is used to balance the differences in dimensions. This is a function for predicting maintenance effects.

[0157] Step B7: Output the maintenance effect prediction function To predict the maintenance effect.

[0158] The above algorithm enables a systematic analysis of pavement performance sequences, quantifying maintenance effectiveness from both short-term jumps and long-term degradation perspectives. This method considers anomaly handling, missing value completion, and dynamic changes in time series, providing reliable support for threshold optimization, maintenance strategy evaluation, and decision-making, thus achieving a closed loop between pavement performance prediction and maintenance optimization.

[0159] 3. Two-level optimization decision model

[0160] 3.1 Modeling Approach

[0161] The core objective of preventative road maintenance is to minimize the present value of total life-cycle costs while ensuring service levels. Total costs consist of institutional costs. With user costs constitute:

[0162]

[0163] (1) Decision structure and logical dependency

[0164] This method unifies the "maintenance trigger threshold strategy" and the "maintenance scheduling decision" into a two-layer optimization framework. The outer layer strategy uses a threshold vector. The parameter, referred to as "strategy θ" in this study, consists of components representing the standardized thresholds for the International Roughness Index, rutting depth, and crack damage, used to determine the triggering conditions for preventative maintenance. The inner response variable is the scheduling matrix. This represents the optimal maintenance execution plan obtained through scheduling optimization under a given threshold strategy θ.

[0165] (2) Two-layer optimization logic

[0166] For any given The system simulation model executes two steps: requirement generation and scheduling optimization. First, based on the performance prediction model... When a certain performance indicator deteriorates to a threshold At that time, corresponding maintenance requirements are generated; then, under the constraints of budget, season, and resources, the optimal implementation sequence and combination of measures for this set of requirements are solved to obtain the scheduling solution. .

[0167] Therefore, the problem can be formalized as a strategy-scheduling bilevel optimization problem:

[0168]

[0169] in The cost function is defined in the next section. This structure guarantees logical consistency and variable independence: the outer layer optimizes the strategy θ, and the inner layer performs the solution execution, realizing a closed loop of the entire process from performance prediction to strategy triggering, maintenance scheduling, and cost assessment.

[0170] 3.2 Definition of Cost Function

[0171] (1) Net Present Value Target

[0172] To evaluate the economic viability of different maintenance threshold strategies over the entire life cycle, this study uses Net Present Value (NPV) as the cost minimization objective. The NPV of the total life cycle cost is defined as:

[0173]

[0174] in, Indicates the length of the analysis period; The social discount rate is used to discount future costs to present value. Indicates the first The actual costs incurred in the year. This formula uses a discount factor. Converting future year expenditures into equivalent present value makes different strategies comparable over time, thus serving as the primary economic objective for optimizing outer-layer decisions.

[0175] (2) Institutional costs

[0176]

[0177] Only when Costs are only included when measure m is actually implemented. Cost per unit This corresponds to the amount of work done.

[0178] (3) User costs

[0179]

[0180] Vehicle operating costs (VOC)

[0181]

[0182] This quadratic model reflects the accelerated increase in marginal operating costs caused by deteriorating road conditions, parameters The results were obtained from stratified regression of samples from multi-province highway management systems.

[0183] Construction delay costs

[0184]

[0185] in, For traffic flow, For the duration of construction disruption, Social value per unit of time.

[0186] Accident risk cost

[0187]

[0188] in, For unit accident losses, Based on joint fitting of accident database and performance database (Poisson or negative binomial regression).

[0189] 3.3 Constraints

[0190] (1) Performance constraints

[0191]

[0192] (2) Annual budget constraints

[0193]

[0194] in, This is the annual budget ceiling.

[0195] (3) Seasonal construction constraints

[0196] Due to the limitations of winter temperature and humidity, some maintenance procedures cannot be carried out in winter. Therefore, the following seasonal constraints are set for the scheduling variables:

[0197]

[0198] That is, for all periods of winter. For maintenance operations such as "micro-surfacing" and "fog sealing", the maintenance schedule must be 0 (construction is not allowed).

[0199] (4) Construction period and traffic organization constraints

[0200]

[0201] in, The maximum available construction period for the quarter. This represents the upper limit on the number of projects that can be run in parallel.

[0202] 3.4 Optimizing Model Representation

[0203] This method formalizes the "threshold strategy-scheduled response" process into a typical simulation-based bi-level optimization model. In this structure:

[0204] The upper layer is responsible for selecting the optimal threshold strategy. And obtain the corresponding execution plan by calling the lower-level scheduling model;

[0205] The lower layer is given Solving for the optimal scheduling under the given conditions And return the cost assessment for use by the upper management.

[0206] (1) Upper layer: Threshold strategy optimization

[0207] The upper-level objective is within the feasible region. The optimal threshold vector is searched internally to minimize the total lifecycle cost.

[0208]

[0209] The upper-level objective function is not in an explicit analytical form, but is a simulated output obtained through lower-level scheduling optimization, thus it belongs to a black-box optimization problem.

[0210] (2) Lower layer: Scheduling optimization based on threshold strategy

[0211] For any candidate threshold strategy provided by the upper layer The objective function for solving the lower-level scheduling optimization problem is:

[0212]

[0213] in, This indicates the optimal maintenance schedule; Indicates at the threshold and maintenance schedule The institutional costs below; Indicates at the threshold and maintenance schedule Lower user costs; This represents the feasible region.

[0214] Among them, feasible region It is determined by the following conditions:

[0215] 1) Based on the maintenance timing threshold Triggered annual maintenance requirements;

[0216] 2) Performance constraints (IRI, Rut, Crack);

[0217] 3) Annual budget constraints;

[0218] 4) Seasonal construction constraints;

[0219] 5) Avoid scheduling feasibility constraints (work period and traffic organization constraints) to avoid conflicts between different work processes.

[0220] (3) Nested solution mechanism

[0221] It is worth noting that the upper layer evaluates each candidate threshold. In each case, the optimal schedule is recalculated through the lower-level model. .Right now:

[0222]

[0223] therefore It is not known in advance, nor does it depend on historical results; instead, it is dynamically generated by the lower-level optimization in each upper-level evaluation. This nested structure ensures the mathematical rigor of the model and the consistency of the policy response.

[0224] 3.5 Solution Algorithm Design

[0225] To efficiently solve the computationally expensive, non-differentiable, and nested two-level simulation optimization problem described above, this method employs a Sequential Surrogate-based Optimization (SSO) framework to implement the two-level optimization decision model. This method iteratively samples and learns within the policy space, using a Gaussian process surrogate model to approximate the costly simulation response. The specific implementation flow of the two-level optimization decision model is as follows.

[0226] Step C1: Input the pavement performance prediction function Maintenance effect prediction function Cost parameters and constraints.

[0227] Step C2: Initialize the design: in the policy space The initial sample set for generating maintenance timing thresholds is generated using Latin hypercube sampling (LHS). ,in This represents the number of samples.

[0228] Step C3: For each sample Proceed to steps C4-C7.

[0229] Step C4: Based on samples and pavement performance prediction function Conduct Monte Carlo simulations to generate maintenance requirements for the planning period, i.e., through... Natural degradation simulation was performed on the pavement performance of this road section. When the pavement performance predicted by the pavement performance prediction model deteriorated to a threshold... When this happens, maintenance is required.

[0230] Step C5: After the maintenance requirement is triggered, with the goal of minimizing the total cost over the entire life cycle, a genetic algorithm is used to solve the scheduling subproblem under the threshold θ, optimizing the implementation sequence and combination of measures for the maintenance requirement to obtain the optimal maintenance schedule. During the solution process, the maintenance effect prediction function is invoked each time a maintenance measure is executed. Predict the effects of maintenance and thus update the future deterioration path of pavement performance.

[0231] Step C6: Calculate the net present value of the total life-cycle cost:

[0232]

[0233] in, It represents the net present value of the total life-cycle cost; Represents the net present value cost function; Indicates at the threshold and maintenance schedule The institutional costs below; Indicates at the threshold and maintenance schedule The user cost is reduced.

[0234] Step C7: If Greater than the set maximum allowable limit Then directly to Imposing extremely high penalties , that is to say This allows the proxy model to identify the region as an infeasible area.

[0235] Step C8: Based on the dataset Training the Gaussian process (GP) surrogate model: .

[0236] Step C9: Continue performing the following operations until the number of iterations reaches the upper limit or the improvement rate falls below the set threshold. Using the Expected Improvement (EI) criterion, in Select the next sampling point and to Perform steps C3-C7 to expand the sample set and refit. Model.

[0237] Step C10: Output the optimal maintenance timing threshold .

[0238] This embodiment also provides a computer device, including: at least one processor, at least one memory, and computer program instructions stored in the memory, which implement the above-described method when executed by the processor.

[0239] This embodiment also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method.

[0240] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0241] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0242] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0243] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0244] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An adaptive optimization decision method for the timing of preventive maintenance of pavements, characterized in that, The method comprises the following steps: S1. Based on historical pavement performance data, traffic load characteristics, climate environment characteristics and material structure characteristics, a pavement performance prediction model is established by using a machine learning algorithm to predict the pavement performance at a future time point by using the pavement performance prediction model; S2. Based on the performance data of the pavement before and after maintenance construction, parameters representing the maintenance effect are extracted, including the instantaneous performance rebound amount after maintenance and the performance decay rate after maintenance, and a maintenance effect prediction model is established to predict the maintenance effect by using the maintenance effect prediction model; S3. Construct a two-level optimization decision model with the goal of minimizing the total cost over the entire life cycle; the upper level uses maintenance timing thresholds. As a decision variable, when the pavement performance predicted by the pavement performance prediction model deteriorates to a threshold... At that time, the corresponding maintenance requirements are generated; the lower layer solves for the corresponding thresholds. The optimal maintenance schedule for this maintenance requirement. In the lower-level solution process, a maintenance effect prediction model is invoked to predict the maintenance effect, thereby updating the future deterioration path of pavement performance; then, an optimization algorithm is used for simulation optimization, and the optimal maintenance timing threshold is obtained through adaptive search. .

2. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 1, characterized in that, The implementation method of the pavement performance prediction model is: Step A1: input historical performance dataset and XGBoost hyperparameters; the historical performance dataset includes pavement performance data, time-varying features and static features of different time points of the ith road section, the time-varying features consist of traffic load features and climate environment features, and the static features are material structure features; Step A2: Construct training samples using a sliding time window. Each training sample contains the time point preceding the i-th road segment. pavement performance data The current time point of the i-th segment Time-varying characteristics The current time point of the i-th segment static features and the current time point of the i-th segment pavement performance data ; Step A3: data preprocessing is performed on the training samples; Step A4: Based on the XGBoost algorithm, the training sample is used to train and optimize the model parameters to minimize the following objective function : wherein, is the number of training samples, is the pavement performance prediction result, is the pavement performance prediction function, which is a regression model trained by the XGBoost algorithm, used to find the mapping relationship between the input features and the output data; Step A5: outputting a pavement performance prediction function to predict the pavement performance at the future time point using the time-varying and static features at the future time point and the pavement performance data before the future time point.

3. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 2, characterized in that, The current time point of the i-th road segment pavement performance data , , , Representing the current time point of the i-th segment respectively The smoothness index, rut depth index, and crack damage index.

4. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 2, characterized in that, In step A3, the data preprocessing of the training samples specifically includes: Outlier processing: Z-score is used to remove outliers in continuous variables, and box plot method is used to remove outliers in classification or proportion variables; Missing value filling: linear interpolation method is used to fill the missing data of time series characteristics, KNN algorithm is used to fill the sequences with large fluctuations, and the mode within the group is used to fill the missing data of structure characteristics; Standardization: all features are uniformly resampled to quarterly resolution, and Min-Max Scaling is used to standardize the data to the range of [0, 1]; After the training of step A4 is completed, the SHAP analysis framework is used to perform explainability analysis on the pavement performance prediction model, the average Shapley value of each feature is calculated, the global contribution of the pavement performance prediction result is quantified, and the feature importance ranking and dependency graph at the road segment level and network level are generated. After the training of step A4 is completed, the SHAP analysis framework is used to perform explainability analysis on the pavement performance prediction model, the average Shapley value of each feature is calculated, the global contribution of the pavement performance prediction result is quantified, and the feature importance ranking and dependency graph at the road segment level and network level are generated.

5. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 1, wherein, Assume the first This section of road in the quarter Maintenance work was carried out, and If maintenance work is completed and pavement performance testing is conducted at the end of the quarter, the pavement performance evolution model is as follows: wherein, is the pavement performance value obtained at the first inspection after the maintenance is completed; is the instantaneous performance rebound after the maintenance is completed, wherein represents the last inspection quarter before the construction, represents the first inspection quarter after the maintenance construction is completed; is the performance decay rate after the maintenance; Thus, the road surface performance time series of the i-th section is obtained .

6. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 1, wherein, The implementation method of the maintenance effect prediction model is: Step B1 : inputting the time series of road surface performance , first detection after maintenance construction completion , measure type , time-varying characteristics and static characteristics ; Step B2: Constructing analysis window ; Step B3: data preprocessing is performed on the input data; Step B4: Calculate the post-maintenance window performance rebound amount at the instant of maintenance completion and the post-maintenance performance decay rate ; Step B5: a feature data set is constructed, wherein The feature data set is ; Step B6: Input feature data }, to train the maintenance effect prediction model to minimize the weighted mean squared error : wherein, is the number of feature data for training, is a coefficient for balancing the dimension difference, is a maintenance effect prediction function; Step B7: Output the maintenance effect prediction function to predict the maintenance effect.

7. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 6, characterized in that, In step B3, the data preprocessing of the input data includes applying Hampel filtering to identify and remove abnormal points to achieve data cleaning, and linear interpolation is applied to fill the missing values; In Step B4, the amount of rebound of the instantaneous performance at the completion of curing is calculated by the following formula : wherein denotes the median operator; In the post-maintenance interval , the robust linear regression fit : 。 8. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 1, wherein, The implementation method of the double-layer optimization decision model is: Step C1 : inputting a road surface performance prediction function , a maintenance effect prediction function , cost parameters and constraints; Step C2: Initialization of design: Generating initial sample set of maintenance opportunity thresholds using Latin hypercube sampling within the policy space wherein wherein is the number of samples; Step C3: For each sample Steps C4-C7 are performed; Step C4: Based on the sample and the pavement performance prediction function Monte Carlo simulations are performed to generate maintenance needs over the planning horizon, i.e. by performing a natural deterioration simulation of the pavement performance of the road segment, when the pavement performance predicted using the pavement performance prediction model deteriorates to a threshold value, a maintenance need is triggered; Step C5: After triggering the maintenance demand, the scheduling sub-problem under the threshold θ is solved by genetic algorithm to minimize the total life-cycle cost, and the optimal maintenance scheduling is obtained by optimizing the implementation timing and measure combination of the maintenance demand ; in the solving process, every time a maintenance measure is executed, the maintenance effect prediction function is called to predict the maintenance effect and update the future deterioration path of pavement performance; Step C6: calculate the net present value of the total life cycle cost: wherein, represents the net present value of the total life cycle cost; represents the net present value cost function; represents the agency cost under threshold and maintenance schedule ; represents the user cost under threshold and maintenance schedule ; Step C7: If is greater than a set maximum allowable limit then directly impose a very high penalty cost on i.e. set so that the surrogate model identifies this region as infeasible. Step C8: Based on the dataset , train a Gaussian process proxy model: ; Step C9: Continually perform the following until the iteration count reaches an upper limit or the improvement rate falls below a set threshold : Using the desired improvement criterion, select the next sample point and perform steps C3-C7 on the expanded sample set and refit the model ;​ Step C10: output the optimal maintenance timing threshold .

9. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 8, characterized in that, In step C5, the objective function of the lower layer scheduling optimization solution is represented as: wherein, represents the optimal maintenance schedule; represents the agency cost at a threshold and maintenance schedule ; represents the user cost at a threshold and maintenance schedule ; represents the feasible region.

10. The method for adaptive optimization decision of pavement preventive maintenance timing according to claim 9, wherein, feasible region by a maintenance opportunity threshold triggered annual maintenance needs and constraints; the constraints include: 1) performance constraint, the expression is: 2) annual budget constraint, the expression is: wherein is the annual budget cap; 3) seasonal construction constraint, the expression is: i.e. for all in winter time period "Micro-surfacing" and "Fog Seal" maintenance operations, maintenance schedule must be 0, i.e. no construction allowed; 4) construction period and traffic organization constraint, the expression is: wherein, is the maximum available duration of a quarter, is the upper limit on the number of parallel projects.

Images