Road surface disease maintenance decision-making method
By applying reinforcement learning model and Transformer-LSTM model in road maintenance decision-making, and combining monthly road data for action selection and cost evaluation, the problems of data lag and unscientific decision-making in the existing technology are solved, and more efficient and accurate road maintenance decisions are achieved.
Patent Information
- Application Number
- CN202510383309.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-06-27
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology has data lag and a lack of scientific and objective decision-making system in road maintenance decision-making, resulting in incomplete scope of pre-maintenance and excessively extensive methods.
The reinforcement learning model is used combined with the Transformer-LSTM model, and by obtaining monthly data of multiple sub-sections of the road surface, action selection and preventive maintenance cost assessment are performed, and maintenance decisions are optimized to minimize the sum of rewards in each month.
It effectively solves the problem of annual data lag, improves the accuracy of preventive maintenance, and realizes a sustainable road maintenance system by reducing maintenance costs and maintaining road usage performance.
Smart Images
Figure CN120218908A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of road maintenance, and particularly relates to a road surface disease maintenance decision-making method. Background Art
[0002] Due to the continuous influence of traffic loads and the action of the external natural environment, various surface damage diseases appear on the road surface. If timely maintenance and treatment are not carried out, with the continuous development of surface diseases, serious road damage will eventually be formed, affecting the normal service function of the road, causing greater maintenance costs, energy resource consumption, and generating a large amount of greenhouse gas emissions. Therefore, preventive maintenance of road surface diseases is an important guarantee for maintaining road performance, reducing maintenance costs, and building a green and sustainable road project. However, there have always been problems in road preventive maintenance, such as incomplete pre-maintenance scope and overly extensive pre-maintenance methods. This is mainly due to two reasons: First, there is a lack of detailed and accurate road surface disease and natural environment monitoring data, and even if there is such data, it is impossible to ensure the coherence and accuracy of the data; Second, there is a lack of a scientific and objective pre-maintenance decision-making system, and maintenance mainly relies on probability models and subjective experience.
[0003] Maintenance decision-making first requires timely and accurate data support. The data sources for maintenance decision-making currently mainly utilize the road detection data carried out on the road every year. However, the statistics of annual data have obvious lag, and most road surface diseases have caused serious damage and urgently need maintenance. This leads to the problem that small disease statistics are easily ignored, resulting in incomplete and inaccurate statistics, and it is even more impossible to carry out timely preventive maintenance. In addition, for some local sections, spot checks are carried out during a specific time period, so that special preventive maintenance can be carried out on some road surface damages. However, this kind of maintenance has low efficiency, a small scope of implementation, high requirements for equipment and staff, and has not reached the level of large-scale promotion.
[0004] Over the years, with the continuous development of information technology, especially with the development of deep learning technology, the road maintenance decision-making method has been continuously improved. Road maintenance decision-making methods can be mainly divided into three types: mathematical model-driven methods, mathematical model-experience fusion-driven methods, and data-driven methods. For mathematical model-driven methods, it means using some mathematical formulas obtained through probability-based or physical model derivation to measure the performance changes of roads, and then making decisions based on this. This method is objective and efficient, and can determine very accurate maintenance timing and costs. However, relying solely on theoretical models is likely to deviate greatly from the actual situation. In addition, for the mathematical model-experience fusion-driven method, on the basis of the theoretical model, empirical parameters are added for correction, or mathematical formulas are directly fitted according to historical experience. This method is closer to the real situation in terms of results, but still produces a large deviation over time and has very poor representativeness.
[0005] The data-driven method refers to a method that uses various data collected from the real world as the basis and then uses algorithms or deep learning models to make maintenance decisions. This method has achieved a relatively high improvement in accuracy, significantly enhanced the representativeness and robustness of the method, and has been applied and verified in some projects. Among them, heuristic methods are considered to be the most effective in solving this type of problem because of their higher computational efficiency. The genetic algorithm is one of the most commonly used heuristic methods in the data-driven method for pavement maintenance decision-making. However, the genetic algorithm can only find a solution close to the optimal one and is prone to converge to a local optimum in some cases. Therefore, improved genetic algorithm techniques and other new heuristic algorithms have also been introduced into pavement maintenance decision-making. Recently, some researchers have used reinforcement learning to solve the road maintenance optimization problem. Reinforcement learning is a subfield of machine learning where an intelligent agent learns an optimal policy by interacting with the environment to maximize long-term cumulative rewards. Compared with other optimization algorithms, reinforcement learning learns an optimal policy that maps states to actions, or learns a value function that maps states to the expected return of specific state-action pairs. It makes use of individual behavior interactions to make the search more efficient. Reinforcement learning has also been proven by many studies to provide flexibility in decision-making.
[0006] In the previous process of using the reinforcement learning model to make road surface maintenance decisions, usually the annual time unit is used to study a single lane; due to the low frequency of data collection, this leads to the sparsity of samples, which in turn limits the generalization ability of the reinforcement learning model and reduces the accuracy of maintenance decisions. Summary of the Invention
[0007] Based on this, it is necessary to provide a road surface disease maintenance decision-making method for the above technical problems.
[0008] An embodiment of the present invention provides a road surface disease maintenance decision-making method, including: Obtain the road condition data of multiple sub - sections of the road surface in any month as the initial state; and preset an action space composed of multiple actions, where the actions are road maintenance measures; Based on the initial state, select an action from the action space through a reinforcement learning model, and determine the preventive maintenance cost according to the taken action; determine the reward according to the preventive maintenance cost, the user usage cost caused by not taking any action, and the environmental loss cost caused by not taking any action; Predict the initial state after taking the action through a road surface disease development prediction model to obtain the road condition data of the next month after taking the action; the road surface disease development prediction model is trained by using the historical road condition data of multiple sub - sections of the road surface for a Transformer - LSTM model; Select an action from the action space according to the road condition data of the next month after taking the action, and obtain the road condition data of each month within a preset time period through multiple rounds of iteration. Taking the minimization of the sum of rewards of each month as the optimization goal, train the reinforcement learning model to obtain a maintenance decision - making model; During the road surface disease maintenance decision - making process, input the road condition data of multiple sub - sections of the road surface to be maintained in the current month into the maintenance decision - making model to obtain the best preventive maintenance decision plan.
[0009] Optionally, the action space includes: no operation, seal coat, microsurfacing, in - place heat treatment, pothole patching, thin - layer overlay, and pothole patching combined with thin - layer overlay.
[0010] Optionally, determining the reward according to the preventive maintenance cost, the user usage cost caused by not taking any action, and the environmental loss cost caused by not taking any action specifically includes: Take the negative of the sum of the preventive maintenance cost, the user usage cost caused by not taking any action, and the environmental loss cost caused by not taking any action as the reward, and the formula is expressed as: ; ; ; where, n is the road section, t is the month, is the lane, is the preset time period, S is the set of road sections, is the number of road sections, is the road section n at month t the reward of, is the road section n at montht The sum of the preventive maintenance cost, the user usage cost caused by taking no action, and the environmental loss cost caused by taking no action for the lane on the upper section n in the month t of the preventive maintenance cost for the lane on the upper section n in the month t of the user usage cost caused by taking no action for the lane on the upper section n in the month t of the environmental loss cost caused by taking no action
[0011] Optionally, the Transformer-LSTM model is trained with the historical road condition data of multiple sub-sections of the road surface, specifically including: Construct a Transformer-LSTM model including an improved Transformer module and a long short-term memory network LSTM connected in sequence, where the improved Transformer module is obtained by connecting a second residual connection and a layer normalization module after the feed-forward network of the Transformer module; Obtain the historical road condition data and historical road surface disaster conditions of multiple sub-sections of the road surface, and arrange the historical road condition data of multiple sub-sections of the road surface in months as the time unit to obtain a sequence data set; Input the sequence data set into the Transformer-LSTM model, obtain the attention distribution of the sequence data set in different sub-spaces through the multi-head attention mechanism of the Transformer-LSTM model to obtain multiple semantic associations; perform normalization processing on the multiple semantic associations through the first residual connection and the layer normalization module of the Transformer-LSTM model to obtain a fused feature; perform weight update on the fused feature through the feed-forward network of the Transformer-LSTM model to obtain a low-order feature; perform normalization processing on the low-order feature through the second residual connection and the layer normalization module to obtain a high-order feature; pass the high-order feature through the long short-term memory network LSTM to obtain a time series prediction result representing the development status of the road surface disaster, and train the Transformer-LSTM model with the objective of minimizing the deviation between the historical road surface disaster condition and the time series prediction result to obtain a trained road surface disease development prediction model.
[0012] Optionally, the road condition data includes road surface disease data and road environment data; The road surface disease data includes: transverse cracks, longitudinal cracks, crocodile cracks, potholes, transverse repairs, longitudinal repairs, block repairs, and looseness; The road environment data includes: the temperature, humidity, rainfall, light intensity, and traffic volume at the road section.
[0013] Optionally, the road surface diseases are obtained by quantifying the disease quantity and area in the road disease data collected through the lightweight detection of the road by drones; the road environment data is obtained by using vehicle-mounted sensors to collect real-time environment data through the vehicle networking technology.
[0014] Optionally, it further includes that before inputting the sequence data set into the Transformer-LSTM model, the time units of the road environment data in the sequence data set are unified through data preprocessing; the specific process of the data preprocessing includes: Calculate the daily average value of each month, the average value of the daily maximum value, the average value of the daily minimum value, and the standard variance of the temperature, humidity, and light intensity respectively , and the calculation formula is: For the rainfall and traffic volume, extract the single-day maximum value in each month and the cumulative sum , and the calculation formula is: Among them, represents the average value, N represents the total number, represents any one of the rainfall and traffic volume, represents any one of the temperature, humidity, and light intensity, is the temperature, is the humidity, is the light intensity, is the rainfall, is the traffic volume.
[0015] Optionally, input the road condition data of multiple sub-sections on the surface of the road to be maintained in the current month into the maintenance decision-making model to obtain the best preventive maintenance decision plan, which specifically includes: Take the road condition data of multiple sub-sections on the surface of the road to be maintained in the current month as the current state; According to the current state, select an action from the action space through the reinforcement learning model; The initial state after taking an action is predicted through a road surface disease development prediction model to obtain the road condition data for the next month after taking the action; an action is selected from the action space according to the road condition data for the next month after taking the action, and the optimal preventive maintenance decision plan is obtained through multiple rounds of iteration.
[0016] For the above road surface disease maintenance decision method provided by the embodiments of the present invention, compared with the prior art, its beneficial effects are as follows: Based on the initial state, the present invention selects an action from the action space through a reinforcement learning model, and then predicts the initial state after taking the action through a road surface disease development prediction model to obtain the road condition data for the next month after taking the action; in this process, taking the monthly data as the time scale provides effective data support for the maintenance decision, can solve the problem of annual data lag caused by taking the year as the time scale in the prior art, overcome the limitation of the lack of real data, and thus improve the accuracy of preventive maintenance; In addition, during the training process of the reinforcement learning model of the present invention, the negative sum of the preventive maintenance cost, the user usage cost caused by not taking an action and the environmental loss cost is used as the reward, and minimizing the sum of the rewards for each month is used as the optimization goal, effectively reducing the cost of road maintenance and maintaining the usage performance of the road, which is an important part of building a low-carbon and environmentally friendly sustainable development road maintenance system. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a method flow chart of a road surface disease maintenance decision method provided in an embodiment; Figure 2 It is a model structure diagram of a road surface disease maintenance decision method provided in an embodiment; Figure 3 It is an application process schematic diagram of a road surface disease maintenance decision method provided in an embodiment; Figure 4 It is a schematic diagram of the simulation result of a road surface disease maintenance decision method provided in an embodiment, Figure 4 where (a) is the simulation result of a two-lane section, Figure 4 and (b) is the simulation result of a three-lane section. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0018] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0019] In one embodiment, a road surface disease maintenance decision method is provided, as Figure 1As shown in the figure, the method includes: Obtain the road condition data of multiple sub - sections on the road surface in any month as the initial state; and preset an action space composed of multiple actions, where the actions are road maintenance measures.
[0020] Based on the initial state, select an action from the action space through a reinforcement learning model, and determine the preventive maintenance cost according to the taken action; determine the reward according to the preventive maintenance cost, the user usage cost caused by not taking the action, and the environmental loss cost caused by not taking the action.
[0021] Predict the initial state after taking the action through a road surface disease development prediction model to obtain the road condition data of the next month after taking the action. The road surface disease development prediction model is trained by using the historical road condition data of multiple sub - sections on the road surface for the Transformer - LSTM model.
[0022] Select an action from the action space according to the road condition data of the next month after taking the action, and obtain the road condition data of each month within a preset time period through multiple rounds of iteration. Taking the minimization of the sum of rewards of each month as the optimization goal, train the reinforcement learning model to obtain a maintenance decision - making model.
[0023] In the process of road surface disease maintenance decision - making, input the road condition data of multiple sub - sections on the road surface to be maintained in the current month into the maintenance decision - making model to obtain the best preventive maintenance decision plan.
[0024] Provide a specific embodiment of the present invention, including: Step 1: Build a road intelligent maintenance database with road condition data. The road condition data includes road surface disease data and road environment data. Among them, the road surface disease data includes: transverse cracks, longitudinal cracks, crocodile cracks, potholes, transverse repairs, longitudinal repairs, block repairs, and looseness. The road surface diseases are obtained by quantifying the disease quantity and area in the road disease data collected by the lightweight detection of the road by drones. Through this lightweight road detection technology, efficient and accurate road disease collection can be realized, which is suitable for the daily inspection of roads, and the road disease data statistics can be carried out on a monthly basis for a certain section of the road to establish an intelligent database of road surface diseases.
[0025] The road environment data refers to the temperature, humidity, rainfall, light intensity, and traffic volume of the section. It is obtained by using in - vehicle sensors to collect real - time environment data through vehicle - to - everything technology, realizing high - precision data collection of specific locations. These environmental parameters can be statistically analyzed on a daily basis and summarized into the road intelligent maintenance database.
[0026] The collection of environmental data relies on the rapid development of current intelligent vehicle and Internet of Things technologies. Intelligent vehicles are integrated with a large number of sensors, which can effectively collect environmental data of the surrounding environment during vehicle driving. Then, through Internet of Things technology, the data is transmitted to the cloud platform for unified storage and management. This Internet of Things vehicle-mounted environmental perception system has made substantial progress and is applied in many intelligent vehicles. Utilizing these data to further promote data sharing and establish a data wisdom library is an important link in the sustainable development of road traffic.
[0027] Step 2: Establish a prediction model for the development of road surface diseases based on the Transformer-LSTM model. The specific process is as follows: (1) Data preprocessing.
[0028] First, before inputting the sequence data set into the Transformer-LSTM model, the time units of the road environmental data in the sequence data set are unified through data preprocessing. The specific process of data preprocessing includes: A Calculate the daily average value, the average value of the daily maximum value, the average value of the daily minimum value, and the standard variance of temperature, humidity, and light intensity respectively on a monthly basis. , and the calculation formula is: ; ; ; For rainfall and traffic volume, extract the single-day maximum value and the cumulative sum , and the calculation formula is: ; ; ; Among them, represents the average value, N represents the total number, represents any one of rainfall and traffic volume, represents any one of temperature, humidity, and light intensity, is the temperature, is the humidity, is the light intensity, is the rainfall, is the traffic volume B Summarize these data and the road surface disease data to form a data set. Each set contains all the data conditions of this month. The sets are arranged together according to the time unit to form the sequence data set for training.
[0029] C Then, normalize these data so that they can be analyzed within a certain scale range to avoid the problem of result distortion caused by an overly large value of a single variable.
[0030] (2) Construct a Transformer-LSTM model, and the specific process is as follows: As Figure 2 shown, the Transformer-LSTM model includes an improved Transformer module and a long short-term memory network LSTM connected in sequence. The Transformer module includes a multi-head attention mechanism, a first residual connection, a layer normalization module, and a feed-forward network. The improved Transformer module is obtained by connecting a second residual connection and a layer normalization module after the feed-forward network of the Transformer module.
[0031] Training process of the Transformer-LSTM model: Obtain the historical road condition data and historical road surface disaster conditions of multiple sub-sections of the road surface, and arrange the historical road condition data of multiple sub-sections of the road surface in monthly time units to obtain a sequence data set.
[0032] Input the sequence data set into the Transformer-LSTM model. Through the multi-head attention mechanism of the Transformer-LSTM model, obtain the attention distribution of the sequence data set in different sub-spaces to get multiple semantic associations; normalize the multiple semantic associations through the first residual connection and layer normalization module of the Transformer-LSTM model to obtain a fused feature. Update the weights of the fused feature through the feed-forward network of the Transformer-LSTM model to obtain a low-order feature. Normalize the low-order feature through the second residual connection and layer normalization module to obtain a high-order feature. Pass the time feature to the high-order feature through the long short-term memory network LSTM to obtain a time series prediction result representing the development status of road surface disasters. Take minimizing the deviation between the historical road surface disaster conditions and the time series prediction result as the optimization objective to train the Transformer-LSTM model to obtain a trained prediction model for the development of road surface diseases.
[0033] A Perform simple data cleaning. According to the law of large numbers, eliminate the abnormal data in the prepared data set. Then use the cubic interpolation method to supplement the eliminated data.
[0034] Train the Transformer-LSTM model. The sequence dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. The dataset needs to contain highway data of different road surface materials in different regions to improve the generalization of the model. During training, a time series of the previous year needs to be input into the model, and the model will predict and output the time series for the next year. Train until the loss function decreases and stabilizes, and then end the training.
[0035] (3)Predict the development status of road surface diseases in the next year. Use the trained Transformer-LSTM model. After processing the road surface disease data and environmental data collected in the past year of the section to be maintained according to the operations in step 2(1), input them into the model, and obtain the time series result of the development status of road surface diseases in the next year through model prediction. According to this time series result, the development status data of road surface diseases in the next year can be finally restored.
[0036] The data used for predicting the development status of road surface diseases is the road surface disease data and corresponding environmental data corresponding to each region after dividing the section into regions in advance. The spatial positions of these data should be accurately numbered to facilitate determining the correspondence between the data and the section. However, before training, the order of these data should be shuffled and then trained to avoid contingency during model training.
[0037] Step 3: Establish a multi-lane preventive maintenance cost model. The multi-lane preventive maintenance cost model proposed in the present invention assumes a two-lane or three-lane model composed of sections. The sections are represented by the set S :
[0038] ; where S is the section set, is the number of sections, Step 4: DRL maintenance decision-making algorithm. DRL is the abbreviation of the deep reinforcement learning neural network. There are usually five key components in the reinforcement learning model, namely the agent, the environment, the state, the action, and the reward.
[0039] (1)Agent and environment: For the preventive maintenance of highway road surface diseases, the agent is the decision-maker responsible for formulating preventive plans. After observing the road surface conditions and considering all other factors that may affect preventive maintenance decisions, the decision-maker implements preventive maintenance activities on the road surface. The environment includes the road surface section itself and its external environment. In the present invention, a simulation environment for preventive maintenance is constructed, which includes a set of road surface disease status development models and reward systems predicted by the Transformer-LSTM model in step 2.
[0040] (2) State: In the problem of preventive maintenance of road surface diseases, the state represents the minimum amount of information required for an agent to make preventive maintenance decisions in the simulation environment. The state variables need to cover all or at least the most important factors involved in the pavement performance model, while excluding variables that remain constant between different road sections and throughout the planning scope, or variables that can be fully inferred from other variables.
[0041] During the training process of the reinforcement learning model, the road maintenance measures combinations of no operation, seal coat, microsurfacing, in-place heat treatment, pothole patching, thin layer overlay, pothole patching and thin layer overlay are used as the action space.
[0042] During the application process of the maintenance decision model, the road condition data of multiple sub-sections of the road surface to be maintained in the current month is used as the current state. According to the current state, an action is selected from the action space through the reinforcement learning model, as shown in Table 1. The initial state after taking the action is predicted through the road surface disease development prediction model to obtain the road condition data for the next month after taking the action. An action is selected from the action space according to the road condition data for the next month after taking the action, and the optimal preventive maintenance decision plan is obtained through multiple rounds of iteration.
[0043] Table 1 Model Application Table (3) Action: In the problem of preventive maintenance of road surface diseases, the action space usually refers to a set of available preventive maintenance treatment measures. In this study, the joint preventive maintenance of multiple lanes of a certain road section is regarded as an action set. Let be the month t and n be the lane l of the road section n for preventive maintenance treatment, then the joint preventive maintenance treatment of the road section
[0044] Therefore, the number of joint M&R treatment measures for a one-way three-lane road section can reach 13^3 = 2197.
[0045] (4) Reward: The negative value of the sum of the preventive maintenance cost, the user usage cost caused by not taking an action, and the environmental loss cost caused by not taking an action is used as the reward, and the formula is expressed as: ; ; where n is the road section, t is the month, is a lane, is a preset time period, is a road section n in a month t of the reward, is a road section n in a month t the sum of the preventive maintenance cost, the user usage cost caused by not taking action, and the environmental loss cost caused by not taking action, is a lane on the road section n in a month t of the preventive maintenance cost, is a lane on the road section n in a month t the user usage cost caused by not taking action, is a lane on the road section n in a month t the environmental loss cost caused by not taking action.
[0046] The training process of the reinforcement learning algorithm includes: Obtain the road condition data of multiple sub - road sections on the road surface in any month as the initial state; and preset an action space composed of multiple actions, where the actions are road maintenance measures.
[0047] Based on the initial state, select an action from the action space through the reinforcement learning model, and determine the preventive maintenance cost according to the taken action. Determine the reward according to the preventive maintenance cost, the user usage cost caused by not taking action, and the environmental loss cost caused by not taking action.
[0048] Predict the initial state after taking the action through the road surface disease development prediction model to obtain the road condition data of the next month after taking the action. The road surface disease development prediction model is trained by the historical road condition data of multiple sub - road sections on the road surface for the Transformer - LSTM model.
[0049] Select an action from the action space according to the road condition data of the next month after taking the action, and obtain the road condition data of each month within the preset time period through multiple rounds of iteration. Take minimizing the sum of the rewards of each month as the optimization goal to train the reinforcement learning model to obtain the maintenance decision - making model.
[0050] As Figure 3 shown, the application process of the maintenance decision - making model includes: Month t Initialization; Obtain the road surface disease state; Selection action; Calculate reward; Update the road table status; Judge t Whether it is equal to T : If t Not equal to T , then set t Update to t +1, update the status and return to obtaining the road surface disease status of the road table; If t Equal to T , then merge and output the actions and rewards selected for each month; Among them, T Is the last month in the preset time period τ in.
[0051] Step 5: Preventive maintenance decision-making. Through simulation experiments, the simulation results are as Figure 4 shown. According to the above model, the preventive maintenance process is deduced and simulated, and iterative optimization is carried out through the value iteration method to generate the best preventive maintenance decision plan.
[0052] The beneficial effects that the present invention can achieve: 1. Combining lightweight road detection by drones, vehicle networking technology, and reinforcement learning models effectively reduces the cost of road maintenance and maintains the performance of roads.
[0053] 2. Based on the initial state, actions are selected from the action space through the reinforcement learning model, and then the initial state after taking actions is predicted through the road surface disease development prediction model to obtain the road condition data for the next month after taking actions; in this process, month data is used as the time scale, providing effective data support for maintenance decision-making, being able to solve the problem of annual data lag caused by the annual time scale in the prior art, overcoming the limitation of the lack of real data, and thus improving the accuracy of preventive maintenance.
[0054] 3. During the training process of the reinforcement learning model, the negative sum of the preventive maintenance cost, the user usage cost caused by not taking actions, and the environmental loss cost is used as the reward, with the goal of minimizing the sum of rewards for each month, effectively reducing the cost of road maintenance and maintaining the performance of roads, which is an important part of building a low-carbon and environmentally friendly sustainable development road maintenance system.
[0055] The above embodiments only represent several implementation manners of the present invention, and the description is relatively specific and detailed, but it should not be construed as a limitation to the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.
Claims
1. A road surface disease maintenance decision-making method, characterized in that: include: Obtaining road condition data of multiple sub-sections of the road surface in any month as an initial state; and presetting an action space consisting of a plurality of actions, wherein the actions are road maintenance measures; Based on the initial state, the reinforcement learning model selects actions from the action space and determines the preventive maintenance cost based on the action taken. The reward is determined based on the preventive maintenance cost, the user cost caused by not taking action, and the environmental loss cost caused by not taking action. The initial state after the action is taken is predicted by a road surface disease development prediction model to obtain the road condition data for the next month after the action is taken; the road surface disease development prediction model is obtained by training a Transformer-LSTM model with historical road condition data of multiple sub-sections on the road surface; According to the road condition data of the next month after the action is taken, an action is selected from the action space. Through multiple rounds of iterations, the road condition data of each month in the preset period is obtained. With the minimization of the sum of rewards of each month as the optimization goal, the reinforcement learning model is trained to obtain the maintenance decision model. In the process of road surface damage maintenance decision-making, the road condition data of multiple sub-sections of the road surface to be maintained in the current month are input into the maintenance decision model to obtain the best preventive maintenance decision plan.
2. A road surface damage maintenance decision-making method as claimed in claim 1, characterized in that: The action space includes: no operation, sealing coating, micro-surfacing, in-situ heat treatment, filling and repairing, thin layer covering, and filling and repairing and thin layer covering.
3. A road surface damage maintenance decision-making method as claimed in claim 1, characterized in that: The rewards are determined based on the preventive maintenance costs, the user usage costs caused by not taking action, and the environmental loss costs caused by not taking action, specifically including: The negative of the sum of the preventive maintenance cost, the user usage cost caused by not taking action, and the environmental loss cost caused by not taking action is used as the reward. The formula is expressed as: ； ； ； in, n For road sections, t For the month, For the lane, For the preset time period, S For road segment collection, is the number of road sections, For road section n In the month t Rewards, For road section n In the month t The sum of the preventive maintenance costs, the user usage costs caused by not taking action, and the environmental loss costs caused by not taking action, For the lane Upper section n In the month t The preventive maintenance cost For the lane Upper section n In the month t User usage costs caused by not taking action, For the lane Upper section n In the month t The environmental costs of not taking action.
4. A road surface damage maintenance decision-making method as claimed in claim 1, characterized in that: The training of the Transformer-LSTM model using historical road condition data of multiple sub-segments on the road surface specifically includes: Constructing a Transformer-LSTM model including an improved Transformer module and a long short-term memory network LSTM connected in sequence, wherein the improved Transformer module is obtained by connecting a second residual connection and a layer normalization module after a feedforward network of the Transformer module; Acquire historical road condition data and historical road surface disaster conditions of multiple sub-sections of the road surface, and arrange the historical road condition data of the multiple sub-sections of the road surface by months as time units to obtain a sequence data set; The sequence data set is input into the Transformer-LSTM model, and the attention distribution of the sequence data set in different subspaces is obtained through the multi-head attention mechanism of the Transformer-LSTM model to obtain multiple semantic associations; the multiple semantic associations are normalized through the first residual connection and layer normalization module of the Transformer-LSTM model to obtain fused features; the weights of the fused features are updated through the feedforward network of the Transformer-LSTM model to obtain low-order features; the low-order features are normalized through the second residual connection and layer normalization module to obtain high-order features; the time features of the high-order features are transferred through the long short-term memory network LSTM to obtain the time series prediction results that characterize the development status of road surface disasters, and the Transformer-LSTM model is trained with the optimization goal of minimizing the deviation between the historical road surface disaster status and the time series prediction results to obtain the trained road surface disease development prediction model.
5. A road surface damage maintenance decision-making method as claimed in claim 4, characterized in that: The road condition data includes road surface damage data and road environment data; The road surface damage data include: transverse cracks, longitudinal cracks, cracks, potholes, transverse repairs, longitudinal repairs, block repairs and looseness; The road environment data includes: temperature, humidity, rainfall, light intensity and traffic volume of the road section.
6. A road surface damage maintenance decision-making method as claimed in claim 5, characterized in that: The road surface disease data is collected by drone road lightweight detection, and the number and area of diseases in the road disease data are quantified; the road environment data is obtained by collecting real-time environment data using vehicle-mounted sensors through vehicle networking technology.
7. A road surface damage maintenance decision-making method as claimed in claim 5, characterized in that: The method also includes, before inputting the sequence data set into the Transformer-LSTM model, unifying the road environment data in the sequence data set into a time unit through data preprocessing; the data preprocessing process specifically includes: Calculate the monthly daily average, average of the daily maximum, average of the daily minimum and standard deviation for temperature, humidity and light intensity respectively. , the calculation formula is: ； ； ； For rainfall and traffic volume, extract the maximum value of each day in each month and the cumulative total , the calculation formula is: ； ； ； in, represents the average value, N Total number of representatives, represents either rainfall or traffic volume, Represents any one of temperature, humidity and light intensity, is the temperature, For humidity, Light intensity, is the rainfall, For traffic volume.
8. A road surface damage maintenance decision-making method as claimed in claim 1, characterized in that: The road condition data of multiple sub-sections of the road surface to be maintained in the current month are input into the maintenance decision model to obtain the best preventive maintenance decision plan, which specifically includes: The road condition data of multiple sub-sections of the road surface to be maintained in the current month are used as the current state; According to the current state, the action is selected from the action space through the reinforcement learning model; The road surface disease development prediction model is used to predict the initial state after the action is taken, and the road condition data for the next month after the action is taken is obtained; actions are selected from the action space based on the road condition data for the next month after the action is taken, and the optimal preventive maintenance decision plan is obtained through multiple rounds of iterations.