A road maintenance decision-making system driven by both physical data and

Abstract

This invention provides a road maintenance decision-making system driven by both physical data and a three-stage maintenance process: pile number, road section, and road network. It can provide pile number-level, road section-level, and road network-level maintenance plans with an accuracy of 100 meters. Each level of maintenance plan simultaneously considers road maintenance construction costs, carbon emissions, construction time, and pavement performance, achieving Pareto optimality among these objectives. This ensures that the final maintenance plan at each level achieves optimal pavement performance while saving construction costs, reducing carbon emissions, and shortening construction time, ultimately achieving the best possible maintenance results. This three-stage maintenance plan eliminates the need for PINN predictions in the most time-consuming phase, directly utilizing a combination of different Pareto optimal maintenance plans from the next-level set, saving significant time.

Description

Technical Field

[0001] The present invention relates to the technical field of road maintenance, and in particular to a road maintenance decision-making system driven by both physical data and physical data. Background Art

[0002] Building an increasingly perfect scientific decision-making management and technical system for highway maintenance, and studying the implementation path of scientific maintenance decision-making driven by both physical data and time are not only important requirements of relevant superior departments for maintenance management, but also rigid demands of maintenance management agencies at all levels in the industry. Therefore, it is very important to establish a scientific, reasonable, sustainable and implementable maintenance decision-making model.

[0003] The shortcomings of existing technologies are that traditional road maintenance relies too heavily on the subjective experience of maintenance engineers. Current advanced maintenance techniques lack consideration for real-world physical constraints and fail to comprehensively factor in factors such as cost, carbon emissions, long-term pavement performance, and construction time. Data-based maintenance decisions, especially those based on deep learning, cannot accurately account for the impact of objective factors during the maintenance process, such as traffic, geology, weather, road structure, and maintenance methods, thus impacting the reliability of these decisions. Summary of the Invention

[0004] In light of the shortcomings of existing maintenance technologies, the present invention aims to address the technical problem of providing a road maintenance decision-making system driven by both physical and data. This system fully and effectively considers the impact of traffic, geological, weather, environmental, road structure, and maintenance methods on pavement performance, forming a three-stage maintenance decision-making system based on pile number, road section, and road network. This system can provide pile-level, road section-level, and road network-level maintenance plans with an accuracy of 100 meters.

[0005] In order to solve the above technical problems, the technical solution of the present invention is:

[0006] The present invention provides a road maintenance decision-making system driven by both physical data and physical data, the system comprising:

[0007] The pavement performance prediction module is used to predict the pavement performance in the next year based on the road age, traffic factors, geological factors, weather and environmental factors, road structure, maintenance methods, and the pavement performance values in the past year;

[0008] The pile number-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of each pile number based on the maintenance plan output by the pile number-level optimization module and the pavement performance predicted by the pavement performance prediction module as the pile number-level fitness;

[0009] The pile number optimization module uses a multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the pile number fitness calculated by the pile number fitness calculation module, outputs a maintenance plan, and obtains the pile number Pareto optimal maintenance plan set for each pile number after reaching the number of iterations. The pile number Pareto optimal maintenance plan set includes G optimal maintenance plans;

[0010] A section-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of each section based on the Pareto optimal maintenance plan set of the pile number and the comprehensive pavement performance of each pile number obtained by the pile number-level fitness calculation module, as the section-level fitness;

[0011] The section-level optimization module inputs the Pareto optimal maintenance plan set for each pile number, and uses the multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the section-level fitness calculated by the section-level fitness calculation module, outputs the maintenance plan, and obtains the Pareto optimal maintenance plan set for each section after the number of iterations is reached;

[0012] A road network-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of the road network based on the Pareto optimal maintenance plan set of the road section and the comprehensive pavement performance of each road section obtained by the road section-level fitness calculation module, as the road network-level fitness;

[0013] The road network level optimization module takes the Pareto optimal maintenance plan set of the road section as input, uses the multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the road network level fitness calculated by the road network level fitness calculation module, outputs the maintenance plan, and obtains the road network Pareto optimal maintenance plan set after reaching the number of iterations.

[0014] Furthermore, the pavement performance prediction module uses a physical embedded neural network (PINN) for prediction, which includes an input layer, a hidden layer, an output layer, and a loss term;

[0015] The loss item It consists of four parts and the calculation formula is:

[0016] (19)

[0017] in, Represents all parameters of the physical embedded neural network, is the data loss coefficient, is the data loss term, is the physical loss coefficient, is the physical loss term, is the boundary loss coefficient, is the boundary loss term, is the regularity loss term coefficient, is the regular loss term.

[0018] Furthermore, in the pile number-level fitness calculation module, the calculation formulas for the cost, carbon emission CE, construction time CT, and comprehensive pavement performance CRPS of each pile number are as follows:

[0019] (4)

[0020] (5)

[0021] (6)

[0022] (7)

[0023] in, It represents the cost of each maintenance plan corresponding to each pile number to adopt the g-th maintenance method in the t-th year, t0 represents the starting year, t e Indicates the year of the end; It represents the carbon emissions emitted by each maintenance plan corresponding to each pile number when the g-th maintenance method is adopted in the t-th year; Indicates the time required for each maintenance plan corresponding to each pile number to be constructed using the g-th maintenance method in year t; the value of g is an integer from 1 to 9, representing micro-surfacing, pavement lining, thermal recycling, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and roadbed grouting, respectively. Represents the physical embedded neural network in the input vector The predicted value below.

[0024] Furthermore, in the section-level fitness calculation module, the calculation formulas for the cost CL, carbon emission CEL, comprehensive pavement performance CRPSL and construction time CTL of each section are as follows:

[0025] (26)

[0026] (27)

[0027] (28)

[0028] (29)

[0029] Where n represents the number of pile numbers; h represents the hth hundred-meter pile number in the road section; represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the pile number h; represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the pile number h+1; 、 The optimal maintenance plan for the section from pile number h to pile number h+1 is arrive Additional costs and construction time; 、 、 、 They represent the optimal maintenance plan M of the pile number h in the Pareto optimal maintenance plan set. h Cost, carbon emissions, pavement performance, and construction time.

[0030] Furthermore, the traffic factors include the average annual passenger car traffic volume, passenger-to-freight ratio, average annual truck traffic volume and equivalent load times; the geological factors include base properties, geological structure and soil characteristics; the weather environment factors include rainfall, average annual daily maximum temperature and average annual daily minimum temperature; the road structure includes surface layer material, surface layer thickness, base layer material and base layer thickness; the maintenance methods include micro-surfacing, pavement patching, thermal regeneration, single-layer excavation and patching, single-layer covering, double-layer excavation and patching, double-layer covering, triple-layer excavation and patching and roadbed grouting.

[0031] Furthermore, the multi-objective exponential distribution optimization algorithm in the pile-level optimization module includes the following process: initializing the population P, the number of individuals in the population is K, and each individual It is an N×M maintenance plan matrix, where N is the length of the maintenance year, M is the number of maintenance methods, and each individual in the population P represents the N-year maintenance plan of the pile number, where each individual maintenance plan satisfies the requirement of only using one maintenance method each year, and each row of the maintenance plan matrix satisfies the constraint shown in formula (9):

[0032] (9)

[0033] Initialize the individual according to formula (10):

[0034] (10)

[0035] Among them, rand(1,M) represents randomly selecting an index column from {1,2,…,M} according to the exponential distribution;

[0036] The guided solution is calculated according to formula (11) using the exploration and utilization principle. :

[0037] (11)

[0038] Among them, time represents the number of iterations, 、 、 Represents the three best solutions at the current number of iterations;

[0039] The output results of the pile-level fitness calculation module are used for information feedback. The update logic of each individual maintenance plan is: based on the pile-level fitness, the weighted coefficients of the guided solution and the memoryless solution are obtained to calculate the value of the new individual matrix. , and The maximum value of each row is assigned to 1, and the other values are assigned to 0. Then the next generation of individuals is updated according to formula (12):

[0040] (12)

[0041] in, It represents selecting the index column j with the largest value for the i-th row of the current matrix, and the corresponding position of the index column with the largest value is assigned a value of 1; is the updated individual in the i-th row, j-th column and time+1-th iteration;

[0042] is defined as:

[0043] (13)

[0044] in, is the memoryless solution of the i-th row and j-th column in the time-th generation individual; a and b are weighting coefficients, defined as:

[0045] (14)

[0046] in, is the fitness of the guided solution, represents the fitness of the kth individual, and both are obtained by the stake-level fitness calculation module.

[0047] Furthermore, the pile number-level fitness calculation module, the section-level fitness calculation module, and the road network-level fitness calculation module all use maintenance specification recommendations as restrictive conditions. When the pavement performance value is below 90, micro-surfacing will not be adopted, and when the pavement performance value is below 80, covering measures will not be adopted.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The system of this invention is a three-stage maintenance decision-making system based on pile number, road section, and road network. It can provide pile number-level, road section-level, and road network-level maintenance plans with an accuracy of 100 meters. Each level of maintenance plan simultaneously considers road maintenance construction costs, carbon emissions, construction time, and pavement performance, achieving Pareto optimality among these objectives. The final maintenance plan at each level saves construction costs, reduces construction carbon emissions, and shortens construction time, while ensuring optimal pavement performance and achieving the best maintenance results.

[0050] 2. The present invention repeatedly utilizes a multi-objective optimization algorithm for multi-level optimization, optimizing the pile number, road section, and road network in stages, significantly improving optimization efficiency. The pile number-level optimization module sets the search space of the multi-objective exponential distribution algorithm to an N×M matrix representing multi-year maintenance plans. During population initialization, a constraint of only one maintenance plan per year is added. During the next generation population update, the algorithm is modified to calculate weights based on fitness values and update the maintenance plan matrix. The largest value in each row is assigned a value of 1, while the others are assigned values of 0. The road section-level optimization module and the road network-level optimization module set the search space of the multi-objective exponential distribution algorithm to a vector consisting of the set of Pareto-optimal maintenance plans for each pile number or the set of Pareto-optimal maintenance plans for each road section. During population initialization and the next generation population update, individual vectors in the population are maintained as integer vectors. This fully accounts for the constraints of actual maintenance operations (each 100-meter pile number is maintained only once per year). The exponential distribution probability allows the algorithm's search space to be efficiently explored, improving optimization efficiency and avoiding local optima.

[0051] 3. This invention uses a PINN (Physical Inline Network) to integrate physical constraints. In the pavement performance prediction module, the invention transforms the impact mechanisms of factors such as road age, traffic factors (average annual passenger vehicle traffic volume, passenger-to-freight ratio, average annual truck traffic volume, and equivalent load times), geological factors (base layer properties, geological structure, soil characteristics), weather and environmental factors (rainfall, average annual daily maximum temperature, average annual daily minimum temperature), road structure (surface material, surface layer thickness, base layer material, base layer thickness), and maintenance methods (micro-surfacing, pavement patching, thermal recycling, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and roadbed grouting) on pavement performance and maintenance effectiveness into partial differential equations. These partial differential equations are then used as loss term constraints in the physical embedded neural network (PINN). The physical embedded neural network (PINN) is trained using these factors, combined with pavement performance values from the past year, as input and pavement performance values from the next year as output. In the fitness calculation, the additional construction time and construction cost required for switching maintenance methods between adjacent 100-meter pile numbers during the maintenance process were taken into account, as well as the maintenance specification recommendations. Both were used as constraints. At the same time, the upper limit of cost, carbon emissions, and construction time during the construction process, as well as the lower limit of pavement performance, were considered. The physical constraints were effectively integrated into a whole, improving the accuracy and stability of prediction and decision-making.

[0052] 4. The three-stage maintenance plan of the present invention eliminates the need for PINN predictions in the most time-consuming phase, directly utilizing combinations of different maintenance plans from the Pareto-optimal maintenance plan set provided by the next level (each maintenance plan combination corresponds to a different fitness, thus directly permuting and combining the next-level maintenance plan strategies without the complex PINN predictions), saving a significant amount of time. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a flow chart of a multi-objective exponential distribution optimization algorithm MOEDO according to an embodiment of the present invention.

[0054] Figure 2 Schematic diagram of the structure of a physical embedded neural network according to an embodiment of the present invention.

[0055] Figure 3 This is a schematic structural diagram of a three-stage maintenance decision system of pile number-road section-road network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following are specific embodiments and drawings of the present invention, which are only used to further illustrate the present invention, but do not limit the scope of protection of this application. It should be noted that the embodiments and features in the embodiments of the present invention can be combined with each other if there is no conflict.

[0057] A road segment is a route between two adjacent nodes on a transportation network. A road network refers to a network of interconnected roads within a certain area. This paper constructs a three-stage maintenance decision-making system: stake number, road segment, and road network. This system can calculate maintenance plans for road segments and even for the road network at the 100-meter stake level, avoiding the problem of being stuck in a local optimal solution due to the large amount of calculation required.

[0058] This invention fully and effectively considers the impact of factors such as road age, traffic factors (average annual passenger vehicle traffic volume, passenger-to-freight ratio, average annual truck traffic volume, and equivalent load times), geological factors (base layer properties, geological structure, and soil characteristics), weather and environmental factors (precipitation, average annual daily maximum temperature, and average annual daily minimum temperature), road structure (surface material, surface layer thickness, base layer material, and base layer thickness), and maintenance methods (micro-surfacing, pavement patching, thermal recycling, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and subgrade grouting) on pavement performance. This invention leverages the multi-objective optimization algorithm MOEDO, combining maintenance requirements at the stake, section, and network levels. By employing a non-dominated sorting and congestion distance mechanism, it effectively balances conflicts among multiple objectives, while also offering the advantages of rapid convergence and a low risk of falling into local optimal solutions. Furthermore, by incorporating the constraints recommended by maintenance specifications in real-world maintenance projects as optimization conditions, it provides a more reliable and feasible maintenance method and offers a more instructive maintenance plan for maintenance projects. The fitness calculation simultaneously considers the cost, carbon emissions, construction time, and overall pavement performance of different levels of road maintenance and construction. This approach aims to achieve optimal overall pavement performance while saving construction costs, reducing carbon emissions, and shortening construction time. This approach achieves the Pareto optimality among these objectives and achieves excellent maintenance results. The maintenance effect refers to the change in pavement performance after maintenance. A sustained high level of performance after maintenance indicates a good maintenance effect, while a rapid decline in performance indicates a poor maintenance effect.

[0059] In the embodiment of the present invention, a physical embedded neural network PINN is constructed for prediction, which fully considers the influence of the above-mentioned influencing factors on pavement performance and maintenance effect, converts them into partial differential equations, and uses the partial differential equations as constraints of the physical embedded neural network to train a pavement performance prediction model that conforms to the real physical laws.

[0060] The present invention includes three modules: an optimization module, a pavement performance prediction module and a fitness calculation module. The optimization module is divided into a pile number level optimization module, a section level optimization module and a road network level optimization module. Each optimization module uses a multi-objective optimization algorithm to output a Pareto optimal maintenance plan set of each level. The fitness calculation module is divided into a pile number level fitness calculation module, a section level fitness calculation module and a road network level fitness calculation module. The cost, carbon emissions, construction time and comprehensive pavement performance are calculated according to the maintenance plans of different levels, and the four are used as fitness. The pavement performance prediction module is used to predict the pavement performance in the next year based on the road age, traffic factors, geological factors, weather and environmental factors, road structure, maintenance methods and the pavement performance values of the past year. Among them, the optimization module adopts a multi-objective exponential distribution optimization algorithm, and the exponential distribution is used in operations including population initialization, exploration, utilization and population update. Figure 1 This is a flowchart of the multi-objective exponential distribution optimization algorithm. The specific process is: first, initialize the population according to the exponential distribution, initialize the parameters to create an initialization population containing K random solutions, generate a child population from the parent population, set k=1, use the information feedback mechanism IFM to update individuals, add the updated individuals to the population until K individuals are obtained, merge the parent and child populations, perform non-dominated sorting and divide the non-dominated levels, select individuals with high dominance levels, and complete the population number according to the crowding ranking. Check whether the number of iterations has been reached. If so, end the iteration and output the result to obtain the Pareto optimal solution. If not, continue the iteration.

[0061] The exponential distribution function f(x) in the present invention follows the following formula:

[0062] (1)

[0063] in, is the exponential distribution parameter; x represents the input;

[0064] but The cumulative distribution function F(x) is:

[0065] (2)

[0066] In order to prevent the algorithm from falling into the local optimal solution, the memoryless probability distribution P is introduced. The specific formula is as follows:

[0067] (3)

[0068] Here, s represents the time at which an event has already occurred, and t represents the length of time that has passed since time s. This formula indicates that the probability of the future is only related to the current time period t, and has nothing to do with the time s that has already passed.

[0069] Example 1:

[0070] In this embodiment, the pile number-level optimization module takes as input the pile number-level fitness corresponding to the pile number and outputs the maintenance plan for each pile number. After the iteration is complete, a Pareto-optimal maintenance plan set is output. This Pareto-optimal maintenance plan set contains multiple optimized maintenance plans, i.e., multiple optimal maintenance plans. The pile number-level fitness includes the cost, carbon emissions, construction time, and overall pavement performance of each pile number, which are obtained by the pile number-level fitness calculation module.

[0071] In the pile number level fitness calculation module, the calculation formulas for the cost, carbon emission CE, construction time CT and comprehensive pavement performance CRPS of each pile number are as follows:

[0072] (4)

[0073] (5)

[0074] (6)

[0075] (7)

[0076] in, It represents the cost of each maintenance plan corresponding to each pile number to adopt the g-th maintenance method in the t-th year, t0 represents the starting year, t e Indicates the year of the end; It represents the carbon emissions emitted by each maintenance plan corresponding to each pile number when the g-th maintenance method is adopted in the t-th year; Indicates the time required for each maintenance plan corresponding to each pile number to be constructed using the g-th maintenance method in year t; the value of g is an integer from 1 to 9, representing the nine maintenance methods: micro-surfacing, pavement lining, thermal recycling, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and roadbed grouting. Represents the physical embedded neural network in the input vector The predicted value under the condition of pavement performance prediction module is the output of the pavement performance prediction module; the input vector It refers to the road age, traffic factors, geological factors, weather and environmental factors, road structure, maintenance methods and road performance values in the past year.

[0077] The pile number level optimization module uses a multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the pile number level fitness calculated by the pile number level fitness calculation module, outputs a maintenance plan, and obtains the pile number Pareto optimal maintenance plan set for each pile number after reaching the number of iterations. The pile number Pareto optimal maintenance plan set includes G optimal maintenance plans; the number of optimal maintenance plans G contained in the pile number Pareto optimal maintenance plan set is determined by the non-dominated sorting and congestion selection results. Each optimal maintenance plan is calculated separately when performing fitness calculation.

[0078] The multi-objective exponential distribution optimization algorithm in the pile-level optimization module first initializes the population according to the exponential distribution, generates a new generation of population according to formula (8), merges the parent generation and the offspring generation, performs non-dominated sorting and divides the non-dominated level, selects individuals with high dominance level, and completes the population according to the crowding ranking. It checks whether the number of iterations has been reached. If so, the iteration ends and the result is output. If not, the iteration continues. It includes the following process:

[0079] Initialize the population P, the number of individuals in the population is K, and each individual It is an N×M matrix, where N represents the duration of the maintenance year and M represents the number of maintenance methods. The initial population is expressed as follows using formula (8):

[0080] (8)

[0081] Each individual in population P represents the N-year maintenance plan of the pile number, where each individual maintenance plan needs to satisfy the requirement of only using one maintenance method each year. Each individual represents a solution, that is, each row of the matrix satisfies the constraint shown in formula (9):

[0082] (9)

[0083] Among them, M is the type of maintenance method, represents the value of the kth individual in the i-th row and j-th column. When each individual in the population is initialized, it is still necessary to meet the requirement of only using one maintenance method each year, and initialize it according to the following formula:

[0084] (10)

[0085] Here, rand(1,M) represents randomly selecting an index column from {1,2,…,M} according to an exponential distribution.

[0086] The exploration and utilization principle is used to calculate the guided solution (current optimal solution) according to formula (11): :

[0087] (11)

[0088] Among them, time represents the number of iterations, 、 、 Represents the three best solutions under the current number of iterations;

[0089] The output of the pile-level fitness calculation module is used for information feedback, which is called the information feedback mechanism (IFM). The update logic of each individual maintenance plan is: the pile-level fitness of the pile-level fitness calculation module is used as the weighted coefficient of the guide solution and the memoryless solution to calculate the value of the new individual matrix. , and assign the maximum value of each row to 1 and other values to 0 to ensure that the constraint of only using one maintenance method each year is met. The next generation of individuals is updated according to the following formula:

[0090] (12)

[0091] Among them, It represents selecting the index column j with the largest value for the i-th row of the current matrix, and the corresponding position of the index column with the largest value is assigned a value of 1; is the updated individual in the i-th row, j-th column and time+1-th iteration;

[0092] is defined as follows:

[0093] (13)

[0094] in is the memoryless solution of the i-th row and j-th column in the time-th generation individual; a and b are weighting coefficients, defined as:

[0095] (14)

[0096] in is the fitness of the guided solution, Represents the fitness of the kth individual. Both fitnesses are obtained by the pile number-level fitness calculation module. There are four fitness values, namely cost, carbon emission, construction time and comprehensive pavement performance. Their respective weighted coefficients are obtained for multi-objective optimization.

[0097] Example 2:

[0098] In this embodiment, the segment-level optimization module and the network-level optimization module use similar multi-objective exponential distribution optimization algorithms for optimization. The input is the Pareto optimal maintenance plan set output by the next-level optimization module, and the output is the Pareto optimal maintenance plan set for each segment or the Pareto optimal maintenance plan set for the network. The specific process is:

[0099] Assume the size of the initial population is K, Represents an n-dimensional integer vector, where n represents the number of hundred-meter stakes that make up a road segment or the number of road segments that make up a road network:

[0100] (15)

[0101] Each individual in the population is initialized according to the following formula, which means that in the range Generate a random integer randomly according to the exponential distribution:

[0102] (16)

[0103] in They represent the upper and lower bounds of the h-th variable, respectively. h represents the h-th hundred-meter stake number that constitutes a road segment or the h-th road segment that constitutes a road network. randint represents the random exponential distribution function.

[0104] The update logic of the individual maintenance plan is: if the current individual is a memoryless solution individual, then update it according to the guided solution and the memoryless solution; if the current solution is not a memoryless solution and the random number d is less than 0.5, then update it according to the memoryless solution and the current solution; otherwise, update it according to the average of all solutions and the current solution. The formula is:

[0105] (17)

[0106] Among them, the values of a and b are the same as those in formula (14), and they are obtained by using the corresponding fitness calculated by the road segment level fitness calculation module or the road network level fitness calculation module; round() represents rounding up to an integer value; , is a random value of (0,1), time is the current number of iterations, Max_time is the maximum number of iterations, is the average of all solutions in the current generation, is the guiding solution of the current generation, is the memoryless solution of the k-th individual in the time-th generation, Substitute the k-th solution for time. is the variance, is the exponential distribution parameter, is a random parameter greater than 1, and d is a random parameter between 0 and 1; is the kth updated individual in the time+1th generation. 、 is defined as:

[0107] (18)

[0108] and Plots the average of all solutions and two individuals randomly selected from the initial population and The distance between them.

[0109] The multi-objective exponential distribution optimization algorithm of this embodiment is as follows: initialize parameters, initialize the initial value of the population according to the exponential distribution, generate the offspring population according to formula (15), use the information feedback mechanism to update the individual values until all individuals in the population are updated, merge the parent and offspring, perform non-dominated sorting and divide the non-dominated levels, select individuals with high dominance levels, complete the population number according to the crowding ranking, check whether the number of iterations has been reached, and if so, end the iteration and output the result; if not, continue the iteration.

[0110] Example 3:

[0111] The pavement performance prediction module in this embodiment uses a physical embedded neural network. Figure 2 This is a schematic diagram of the physical embedded neural network architecture, consisting of four main components: input layer, hidden layer, output layer, and loss term. The input layer includes road age, traffic factors (average annual passenger vehicle traffic volume, passenger-to-freight ratio, average annual truck traffic volume, and equivalent load times), geological factors (base layer properties, geological structure, and soil characteristics), weather and environmental factors (rainfall, average annual daily maximum temperature, average annual daily minimum temperature), road structure (surface material, surface layer thickness, base layer material, and base layer thickness), maintenance methods (micro-surfacing, pavement patching, thermal recycling, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and subgrade grouting), and pavement performance over the past year. The output value represents pavement performance over the next year. The input vector of each layer in the physical embedded neural network is derived from the output vector of the previous layer. The input vector is processed using weights and biases to form the output vector.

[0112] The loss item It consists of four parts and the calculation formula is:

[0113] (19)

[0114] in, Represents all parameters of the physical embedded neural network, is the data loss coefficient, is the data loss term, is the physical loss coefficient, is the physical loss term, is the boundary loss coefficient, is the boundary loss term, is the regularity loss term coefficient, is the regular loss term.

[0115] The data loss term is the error between the predicted value and the true value and is calculated as follows:

[0116] (20)

[0117] in, represents the physical embedded neural network prediction value of the s-th sample, Represents the true value of the data item of the sth sample, represents the number of samples used to calculate the data item, Input vector representing the sth sample data item.

[0118] The physical loss term is calculated according to the following formula:

[0119] (twenty one)

[0120] in, represents the predicted value of the physical item of the sth sample, represents the true value of the physical item of the sth sample, represents the number of samples used to calculate the physical term, Input vector representing the sth sample physical term. The calculation formula is as follows:

[0121] (twenty two)

[0122] in , ,…, is an undetermined parameter, t represents time, is the input variable, n represents the number of input variables; PPI is the pavement performance, Represents the physical embedded neural network as a whole.

[0123] The calculation formula of regularity loss term is as follows:

[0124] (twenty three)

[0125] in, Represents the number of samples used to calculate the regularity loss, represents the input vector of the s-th sample regularity term, Represents the partial derivative of the physical embedded neural network of the sth sample with respect to time t. The regularity of this loss term means that the road performance degrades over time.

[0126] The boundary loss term is calculated as follows:

[0127] (twenty four)

[0128] in, represents the number of samples used to calculate the boundary loss, Represents the output value of the physical embedded neural network when the input vector is 0 vector, Represents the partial derivative of the physical embedded neural network with respect to time t when the input vector is 0.

[0129] Example 4:

[0130] This embodiment also includes a data preprocessing module, and the data preprocessing process is as follows:

[0131] (1) The input data is processed in time series. Each pile number and its corresponding data are arranged vertically according to time. The influencing factor data of each pile number are arranged horizontally. The last column is the pavement performance value.

[0132] (2) The data is processed in a supervised manner, and the next row of data of each row is arranged horizontally behind the row, so that each row contains the influencing factors and pavement performance data of the pile number in the current year and the influencing factors and pavement performance data of the next year.

[0133] (3) Clean out abnormal data, align the pile number, year, and lane of each row of data, and at the same time remove data where the maintenance factor is empty but the pavement performance in the next year is higher than that in the current year, and data where the maintenance factor is not empty but the pavement performance in the next year is lower than that in the current year.

[0134] (4) Use one-hot encoding to convert the maintenance method data into a vector of 0 and 1. The nine maintenance methods are encoded as follows: micro-surfacing [1,0,0,0,0,0,0,0,0], pavement lining [0,1,0,0,0,0,0,0,0], thermal regeneration [0,0,1,0,0,0,0,0,0], single-layer excavation and patching [0,0,0,1,0,0,0,0,0], single-layer overlay [0,0,0,0,1,0,0,0,0], double-layer excavation and patching [0,0,0,0,0,1,0,0,0], double-layer overlay [0,0,0,0,0,0,1,0,0], triple-layer excavation and patching [0,0,0,0,0,0,0,1,0], and roadbed grouting [0,0,0,0,0,0,0,0,1]. The subsequent maintenance plan matrix is also processed according to this coding method.

[0135] (5) Divide the data by time and normalize the data. Divide the data by time into the first N-1 years as the training set data and the last year as the test set data. The purpose of data normalization is to map all data to between 0 and 1 so that the neural network can learn the data features. The specific data normalization formula is as follows:

[0136] (25)

[0137] in, is the maximum value of the current column, is the minimum value of the current column, and X is the data of the current column; is the value after data normalization.

[0138] After the above data preprocessing, a data set is formed, which is used to train the physical embedded neural network architecture of the road performance prediction module. The neural network training process is:

[0139] (1) This embodiment adopts the physical embedded neural network of embodiment 3 and establishes the physical embedded neural network architecture using Python language.

[0140] (2) Set the hyperparameters of the physical embedded neural network. Set the hidden layer to three fully connected layers, set the number of physical embedded neural network units in each hidden layer to 64, set the activation function to the tanh function, set the batch size to 128, set the learning rate to 0.001, set the number of training times to 2000, set the Adam automatic differentiation calculation, and set the loss function to formula (19).

[0141] (3) Set the last column of the data as the output value and the other columns as the input value, and call the fit function of the keras library for training.

[0142] (4) Call the save function in the keras library to save the model parameters.

[0143] When making predictions, ensure that the input data composition is consistent with the training data, and call the load_model function in the keras library to make predictions.

[0144] After the training is completed, a trained model is obtained, and the trained model is used to predict the pavement performance, and the predicted value is used as the basis for the pile number level fitness calculation module.

[0145] Example 5:

[0146] In this embodiment, the structure of the road segment-level fitness calculation module is similar to that of the road network-level fitness calculation module. The road segment-level fitness calculation module is used as an example for explanation:

[0147] In the road segment-level fitness calculation module, the calculation formulas for the cost CL, carbon emission CEL, comprehensive pavement performance CRPSL and construction time CTL of each road segment are as follows:

[0148] (26);

[0149] (27)

[0150] (28)

[0151] (29)

[0152] Where n represents the number of pile numbers; h represents the hth hundred-meter pile number in the road section; represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the pile number h; represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the pile number h+1; 、 The optimal maintenance plan for the section from pile number h to pile number h+1 is arrive Additional costs and construction time; 、 、 、 They represent the optimal maintenance plan M of the pile number h in the Pareto optimal maintenance plan set. h Cost, carbon emissions, pavement performance, and construction time.

[0153] In the road network level fitness calculation module, n represents the number of road segments; h represents the hth road segment in the road network; An optimal maintenance plan in the Pareto optimal maintenance plan set for the hth road section; represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the h+1th road section; 、 They represent the optimal maintenance plan for switching from section h to section h+1. arrive Additional costs and construction time; 、 、 、 They represent the optimal maintenance plan M of the Pareto optimal maintenance plan set for road section h. h Cost, carbon emissions, pavement performance, and construction time.

[0154] In the pile-level fitness calculation module, the maintenance specifications recommend that micro-surfacing be discouraged when the predicted pavement performance value is below 90, and that overlay measures (in this case, single-layer or double-layer overlay) be discouraged when the predicted pavement performance value is below 80. The maintenance plan matrix is composed using the aforementioned one-hot code. The elements in the matrix are represented in order: micro-surfacing, pavement patching, thermal regeneration, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and roadbed grouting. Therefore, this embodiment adopts the following restrictions in the pile-level fitness calculation module based on the maintenance specifications:

[0155] (30)

[0156] in Indicates that the micro-surfacing (first type) maintenance method is not adopted in year t. Similarly, Indicates that the single-layer covering (fourth type) maintenance method is not adopted in year t. It means that the double-layer covering (seventh type) maintenance method will not be adopted in year t.

[0157] Similarly, in the road segment-level fitness calculation module and the road network-level fitness calculation module, the comprehensive pavement performance of each pile number and the comprehensive pavement performance of each road segment also meet the restrictions recommended by the maintenance specifications.

[0158] In addition, the present invention stipulates that the cost, carbon emissions, and construction time of the maintenance plan have upper limits, and the comprehensive pavement performance has lower limits, as shown in the following formula:

[0159] (31)

[0160] in, represents the cost ceiling, Represents the upper limit of construction time, represents the carbon emission cap, Represents the lower limit of comprehensive performance.

[0161] When the optimal maintenance plan provided by the optimization module does not meet the above restrictions, the fitness will provide an infinite value to indicate that the maintenance plan is unreasonable.

[0162] Example 6:

[0163] The specific process of the physical data dual-driven road maintenance decision system in this embodiment is as follows: Figure 3 As shown:

[0164] Step 1. Preprocess the data and train the physical embedded neural network to obtain the pavement performance prediction module.

[0165] Step 2. Enter the pre-maintenance pile number, road section, and road network information. Enter the pre-maintenance pile number, the pre-maintenance year, and the pavement performance data for the lane pile location of the road section and road network in pile numbers.

[0166] Step 3. Pile maintenance decision stage:

[0167] The maintenance plan for each pile number is optimized. (A maintenance plan represents the maintenance method to be adopted each year over N years. The maintenance plan matrix is a digital representation of the maintenance plan. If no maintenance is performed in the first year, the first row is filled with all zeros. If the first maintenance method is adopted in the first year, the first column of the first row is filled with 1, and the remaining columns are filled with 0s. This continues in this manner, forming a maintenance plan matrix consisting of N rows and M columns of 0s and 1s.) The pile number-level optimization module optimizes the maintenance plan and inputs it into the pavement performance prediction module and the pile number-level fitness calculation module. The pavement performance prediction module outputs a predicted pavement performance value based on the influencing factor data that matches the maintenance section location and year information.

[0168] The pile-level fitness calculation module calculates the cost, carbon emissions, construction time, and overall pavement performance of the pile under the maintenance plan provided by the pile-level optimization module and the pavement performance prediction module. This calculation is used as the pile-level fitness and returned to the pile-level optimization module. The pile-level optimization module then performs non-dominated sorting and congestion sorting on the population based on the pile-level fitness, selects a new generation of individuals (each corresponding to a maintenance plan), and provides these maintenance plans to the pile-level fitness calculation module and the pavement performance prediction module. This process is repeated until the iteration termination condition is met, and the Pareto-optimal maintenance plan set for each pile is output. This Pareto-optimal maintenance plan set for each pile is then input into the segment-level optimization module in the segment maintenance decision phase.

[0169] Step 4. Road maintenance decision-making stage:

[0170] Maintenance plans for each road section are optimized. The section-level optimization module uses the Pareto-optimal maintenance plan set for each pile number in the entire section as a population of individuals. After initializing the maintenance plan for each section, it provides it to the section fitness calculation module. The module then calculates the cost, carbon emissions, construction time, and comprehensive pavement performance of the section based on the maintenance plan. This is used as the section-level fitness and returned to the section-level optimization module. The module then sorts the population by non-dominated and congestion scores based on the fitness, selects a new generation of individuals (each corresponding to a maintenance plan), and provides these maintenance plans to the section-level fitness calculation module. This process is repeated until the iteration termination condition is met, outputting the Pareto-optimal maintenance plan set for each section. This Pareto-optimal maintenance plan set for each section is then input into the network-level optimization module in the road network maintenance decision phase.

[0171] Step 5. Road network maintenance decision-making stage:

[0172] Optimize the maintenance plan for the entire road network. The network-level optimization module uses the Pareto-optimal maintenance plan set for each road section in the entire network as a population of individuals. After initializing the network maintenance plan, it provides it to the network-level fitness calculation module. The network-level fitness calculation module calculates the network's cost, carbon emissions, construction time, and comprehensive pavement performance based on the maintenance plan. This is used as the network-level fitness and returned to the network-level optimization module. The network-level optimization module then performs non-dominated and congestion sorting on the population based on the network-level fitness, selects a new generation of individuals (each corresponding to a maintenance plan), and provides these maintenance plans to the network fitness calculation module. This process is repeated until the iteration termination condition is met, outputting the Pareto-optimal maintenance plan set for the network.

[0173] The above-mentioned iteration stopping condition may be whether the set maximum number of iterations is reached.

[0174] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A road maintenance decision-making system driven by both physical data and physical data, characterized by: The system comprises: The pavement performance prediction module is used to predict the pavement performance in the next year based on the road age, traffic factors, geological factors, weather and environmental factors, road structure, maintenance methods, and the pavement performance values in the past year; The pavement performance prediction module uses a physical embedded neural network PINN for prediction, including an input layer, a hidden layer, an output layer and a loss term; The loss term L(θ) consists of four parts, and the calculation formula is: Among them, θ represents all parameters of the physical embedded neural network, w u is the data loss coefficient, is the data loss term, w f is the physical loss coefficient, L PDE (θ) is the physical loss term, w b is the boundary loss term coefficient, Lb(θ) is the boundary loss term, w r is the regular loss term coefficient, L r (θ) is the regularity loss term; The data loss term is the error between the predicted value and the true value and is calculated as follows: in, represents the physical embedded neural network prediction value of the sth sample, u s Represents the true value of the data item of the sth sample, N u represents the number of samples used to calculate the data item, The input vector representing the s-th sample data item; The physical loss term is calculated according to formula (21): in, represents the predicted value of the physical item of the sth sample, f s Represents the true value of the physical item of the sth sample, N f represents the number of samples used to calculate the physical term, The input vector representing the physical term of the sth sample; The calculation formula of regular loss term is: Among them, N r Represents the number of samples used to calculate the regularity loss, represents the input vector of the s-th sample regularity term, represents the partial derivative of the physical embedded neural network of the sth sample with respect to time t; the regular meaning of this regular loss term is that the road performance decreases over time; The boundary loss term is calculated as: Among them, N b Represents the number of samples used to calculate the boundary loss, u θ (0) represents the output value of the physical embedded neural network when the input vector is 0 vector, represents the partial derivative of the physical embedded neural network with respect to time t when the input vector is 0; The pile number-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of each pile number based on the maintenance plan output by the pile number-level optimization module and the pavement performance predicted by the pavement performance prediction module as the pile number-level fitness; The pile number optimization module uses a multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the pile number fitness calculated by the pile number fitness calculation module, outputs a maintenance plan, and obtains the pile number Pareto optimal maintenance plan set for each pile number after reaching the number of iterations. The pile number Pareto optimal maintenance plan set includes G optimal maintenance plans; A section-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of each section based on the Pareto optimal maintenance plan set of the pile number and the comprehensive pavement performance of each pile number obtained by the pile number-level fitness calculation module, as the section-level fitness; The section-level optimization module inputs the Pareto optimal maintenance plan set for each pile number, and uses the multi-objective exponential distribution optimization algorithm to perform multi-objective optimization on the section-level fitness calculated by the section-level fitness calculation module, outputs the maintenance plan, and obtains the Pareto optimal maintenance plan set for each section after the number of iterations is reached; A road network-level fitness calculation module is used to calculate the cost, carbon emissions, construction time and comprehensive pavement performance of the road network based on the Pareto optimal maintenance plan set of the road section and the comprehensive pavement performance of each road section obtained by the road section-level fitness calculation module, as the road network-level fitness; The road network optimization module uses the Pareto optimal maintenance plan set of the road section as input, performs multi-objective optimization on the road network fitness calculated by the road network fitness calculation module using a multi-objective exponential distribution optimization algorithm, outputs a maintenance plan, and obtains the road network Pareto optimal maintenance plan set after reaching the number of iterations; The multi-objective exponential distribution optimization algorithm in the pile-level optimization module includes the following process: initializing the population P, the number of individuals in the population is K, and each individual X (k) It is an N×M maintenance plan matrix, where N is the length of the maintenance year, M is the number of maintenance methods, and each individual X in the population P (k) represents the N-year maintenance plan of the pile number, where each individual maintenance plan satisfies the requirement of only using one maintenance method each year, and each row of the maintenance plan matrix satisfies the constraint shown in formula (9): Initialize the individual according to formula (10): Among them, rand(1,M) represents randomly selecting an index column from {1,2,…,M} according to the exponential distribution; Calculate the guide solution Xguide according to formula (11) time : Among them, time represents the number of iterations, Represents the three best solutions at the current number of iterations; The output results of the pile-level fitness calculation module are used for information feedback. The update logic of each individual maintenance plan is: based on the pile-level fitness, the weighted coefficients of the guided solution and the memoryless solution are obtained to calculate the value of the new individual matrix. and will The maximum value of each row is assigned to 1, and the other values are assigned to 0. Then the next generation of individuals is updated according to formula (12): Among them, argmax j It represents selecting the index column j with the largest value for the i-th row of the current matrix, and the corresponding position of the index column with the largest value is assigned a value of 1; is the updated individual in the i-th row, j-th column and time+1-th iteration; is defined as: in, is the memoryless solution of the i-th row and j-th column in the time-th generation individual; a and b are weighting coefficients, defined as: Among them, f guide is the fitness of the guided solution, f k represents the fitness of the kth individual, and both are obtained by the stake-level fitness calculation module; The segment-level optimization module and the network-level optimization module use similar multi-objective exponential distribution optimization algorithms for optimization. The input is the Pareto optimal maintenance plan set output by the next-level optimization module, and the output is the Pareto optimal maintenance plan set for each segment or the Pareto optimal maintenance plan set for the network. The update logic of the individual maintenance plan is as follows: if the current individual is a memoryless solution, it is updated based on the guided solution and the memoryless solution; if the current solution is not a memoryless solution and the random number d is less than 0.5, it is updated based on the memoryless solution and the current solution; otherwise, it is updated based on the average of all solutions and the current solution. The formula is: Among them, the values of a and b are the same as those in formula (14), and they are obtained by using the corresponding fitness calculated by the road segment level fitness calculation module or the road network level fitness calculation module; round() represents rounding up to an integer value; f is a random value (0,1), time is the current number of iterations, Max_time is the maximum number of iterations, M time is the average of all solutions in the current generation, is the guiding solution of the current generation, is the memoryless solution of the k-th individual in the time-th generation, Substitute the k-th solution for time; is the variance, λ is the exponential distribution parameter, is a random parameter greater than 1, and d is a random parameter between 0 and 1; is the kth updated individual of the time+1th generation; The definitions of Z1 and Z2 are: Z1=M time -D1+D2,Z2=M time -D2+D1,D1=M time -X rand1 ,D2=M time -X rand2 (18) D1 and D2 depict the mean value M of all solutions time and two individuals X randomly selected from the initial population rand1 and X rand2 the distance between them; The pile number-level optimization module sets the search space of the multi-objective exponential distribution algorithm to an N×M matrix to represent multi-year maintenance plans. During population initialization, a restriction of only performing maintenance once a year is added. During the update of the next generation population, the algorithm is modified as follows: weights are calculated using fitness values and the maintenance plan matrix is updated. The largest value in each row is selected and assigned a value of 1, and the others are assigned values of 0. The section-level optimization module and the road network-level optimization module set the search space of the multi-objective exponential distribution algorithm to a vector consisting of the pile number Pareto optimal maintenance plan set for each pile number or a vector consisting of the section Pareto optimal maintenance plan set for each section. During population initialization and the update of the next generation population, the individual vectors in the population are kept as integer vectors.

2. The physical data dual-driven road maintenance decision system according to claim 1 is characterized in that: In the pile number level fitness calculation module, the calculation formulas for the cost, carbon emission CE, construction time CT and comprehensive pavement performance CRPS of each pile number are as follows: Among them, c t (g) represents the cost of each maintenance plan corresponding to each pile number to adopt the g-th maintenance method in the tth year, t0 represents the starting year, t e Indicates the year of the end; q t (g) represents the carbon emissions emitted by each maintenance plan corresponding to each pile number when the g-th maintenance method is adopted in the t-th year; h t (g) represents the time required for each maintenance scheme corresponding to each pile number to be constructed using the g-th maintenance method in year t; the value of g is an integer from 1 to 9, representing micro-surfacing, pavement lining, thermal regeneration, single-layer excavation and patching, single-layer overlay, double-layer excavation and patching, double-layer overlay, triple-layer excavation and patching, and roadbed grouting, respectively; u(x(t)) represents the predicted value of the physical embedded neural network under the input vector x(t).

3. The physical data dual-driven road maintenance decision system according to claim 1 is characterized in that: In the section-level fitness calculation module, the calculation formulas for the cost CL, carbon emission CEL, comprehensive pavement performance CRPSL and construction time CTL of each section are as follows: Where n represents the number of pile numbers; h represents the hth hundred-meter pile number in the road section; M h M represents an optimal maintenance solution in the Pareto optimal maintenance solution set for the hth pile number; h+1 represents an optimal maintenance plan in the Pareto optimal maintenance plan set for the pile number h+1; Indicates that the optimal maintenance plan for the section from pile number h to pile number h+1 is switched by M h to M h+1 Additional cost and construction time; Cost h (M h ), CE h (M h ), CRPS h (M h ), CT h (M h ) represent the optimal maintenance plan M of the pile number h in the Pareto optimal maintenance plan set. h Cost, carbon emissions, pavement performance, and construction time.

4. The physical data dual-driven road maintenance decision system according to claim 1 is characterized in that: The traffic factors include the average annual passenger car traffic volume, passenger-to-freight ratio, average annual truck traffic volume and equivalent load times; the geological factors include base properties, geological structure and soil characteristics; the weather environment factors include rainfall, average annual daily maximum temperature and average annual daily minimum temperature; the road structure includes surface layer materials, surface layer thickness, base layer materials and base layer thickness; the maintenance methods include micro-surfacing, pavement patching, thermal regeneration, single-layer excavation and patching, single-layer covering, double-layer excavation and patching, double-layer covering, triple-layer excavation and patching and roadbed grouting.

5. The physical data dual-driven road maintenance decision system according to claim 1 is characterized in that: The pile number-level fitness calculation module, the section-level fitness calculation module, and the road network-level fitness calculation module all use maintenance specification recommendations as restrictive conditions. When the pavement performance value is below 90, micro-surfacing will not be used, and when the pavement performance value is below 80, covering measures will not be used.

Images