CN113177735B - Highway pavement full life cycle quality tracing method based on artificial intelligence - Google Patents

Highway pavement full life cycle quality tracing method based on artificial intelligence Download PDF

Info

Publication number
CN113177735B
CN113177735B CN202110571984.9A CN202110571984A CN113177735B CN 113177735 B CN113177735 B CN 113177735B CN 202110571984 A CN202110571984 A CN 202110571984A CN 113177735 B CN113177735 B CN 113177735B
Authority
CN
China
Prior art keywords
data
value
index
indexes
parameter
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110571984.9A
Other languages
Chinese (zh)
Other versions
CN113177735A (en
Inventor
李家乐
王雪菲
殷国辉
马国伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hebei University of Technology
Original Assignee
Hebei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hebei University of Technology filed Critical Hebei University of Technology
Priority to CN202110571984.9A priority Critical patent/CN113177735B/en
Publication of CN113177735A publication Critical patent/CN113177735A/en
Application granted granted Critical
Publication of CN113177735B publication Critical patent/CN113177735B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/08—Construction

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Game Theory and Decision Science (AREA)
  • Physiology (AREA)
  • Evolutionary Biology (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method for tracing the quality of a road surface in a full life cycle based on artificial intelligence. On the basis of the performance prediction model, maintenance decisions are provided for the expressway in the operation period, and construction period data of the expressway in the construction period can be planned by combining a genetic algorithm to guide construction. The invention takes various index data of the whole life cycle of the road as an investigation object, respectively processes the data according to different attributes of each index, integrates the data into a database, and realizes the whole life cycle tracing of the highway by combining an artificial neural network and a genetic algorithm.

Description

Highway pavement full life cycle quality tracing method based on artificial intelligence
Technical Field
The invention relates to the technical field of road engineering, in particular to a method for establishing a tracing back by utilizing an artificial intelligence algorithm for the full life cycle of an expressway, which can be used for making maintenance decisions of an operation highway and construction schemes of a road under construction.
Background
For the road in the operation period, under the comprehensive action of vehicle load and natural factors, the service performance of the road surface is gradually attenuated along with the increase of the road age. The concept of a road management system originated in canada in the 60's of the 20 th century and functions as a "coordinated, integrated and unified collection of activities related to road planning, design, construction, maintenance, evaluation and research". The object of the road surface management system is to enable a management department to efficiently use resources through such a platform, to provide and maintain a road surface having a sufficient level of service within a predetermined life span with minimum resource consumption. However, the pavement management system has the main point of being on the pavement, emphasizes on the overall benefit evaluation and the method decision preference of the later operation of the pavement from the overall and systematic level, and does not relate to the recording, feedback and even application of a series of data such as raw materials, construction schemes, construction key mechanical indexes and the like. The main work content of the road surface management system is that network level management and project level management are carried out on the existing road network according to a data acquisition system, so that the service level of the whole road is optimal, and the influence and feedback of the construction of a highway in the construction period on the operation period are not highlighted.
The traditional modeling method is applied to the management work of the expressway, and the road use performance is not accurately predicted from the perspective of the full life cycle, so that the prediction model is not comprehensive enough and the application range is low. The road surface roughness index is predicted by using road age and year average freezing indexes (Yamamy M S, Saeed T U, Volovski M, et al, characterization of the performance of the enterprise flexible performances using road surface roughness [ J ]. Journal of Infrastructure Systems,2020,26(2):04020010.), and because the difference of a large number of characteristic values in the construction period is not considered, modeling prediction can be carried out only on each state alone, and the model cannot be expanded to other expressways. The existing specifications and researches mostly guide the construction period from a macroscopic perspective, give an approximate range of index parameters, and qualitatively avoid and suggest the construction behaviors. Zhangyang (Zhang yang. quality control in highway subgrade and bridge engineering construction [ J ]. intelligent city, 2020,6(23):99-100.) gives solutions and methods for some key problems in highway subgrade and bridge engineering construction, and lists concerned points for construction quality control, however, regarding specific details in construction, no relevant description and research is available for index quantification, and the significance for determining key parameter indexes of specific construction of highways is small. In the existing research, the data of the construction period and the data of the construction period are not deeply mined and comprehensively associated, so that the waste of data resources is directly caused, and the construction indexes of the construction period are not accurately and scientifically guided.
Disclosure of Invention
In view of the defects of the current situation, the technical problems to be solved by the invention are as follows: the method is used for tracing the whole life cycle of the expressway road network from the construction period to the operation period, an artificial neural network with characteristic values containing relevant data of the construction period and the operation period is established, and the pavement service performance of the expressway at regular time and fixed point is output. The prediction function of the artificial neural network obtained by training is applied to maintenance decision of the expressway in the operation period, the feedback analysis function of the neural network guides construction parameters of the built road, the construction parameters are close to or reach construction conditions corresponding to the ideal value of the service life of the road, the service life of the road is controlled manually, and the maintenance cost is reduced.
In order to solve the technical problems, the technical scheme of the invention is as follows: a full life cycle quality tracing method for a road surface based on artificial intelligence is designed, powerful guidance can be provided for maintenance of an operation road and construction of a raised road, and the tracing method comprises the following steps:
the method comprises the following steps: and collecting the full life cycle data of the highway and establishing a database.
The highway full life cycle data specifically comprises construction period data, operation period data and use performance data, the construction period data specifically comprises mechanical indexes and material attribute indexes, and the operation period data specifically comprises traffic condition indexes and environmental factor indexes.
One piece of data in the database takes a name of a highway, a pile number of a starting point, a road grade, planned traffic volume and a region as a fixed label, takes full life cycle data of a road surface between one hundred meters of pile numbers on one lane as a characteristic parameter, one value of the characteristic parameter is a characteristic value of one index in the full life cycle data of the road surface in one time period, the characteristic value is a data point, and the characteristic value of one index in one time period is obtained according to the data in the corresponding time period;
the characteristic parameters comprise fixed characteristic parameters and periodic characteristic parameters, the indexes of the construction period data are the fixed characteristic parameters, and the characteristic values are recorded only once in the whole life cycle of the road surface; the cycle characteristic parameters comprise traffic condition indexes of operation period data, environment factor indexes of the operation period data and indexes of using performance data, and the cycle characteristic parameters are respectively recorded in different time periods, namely the recording times of the cycle characteristic parameters are the number of time periods included in the full life cycle.
Step two: and (4) preprocessing a plurality of pieces of data of the database in the step one to obtain an original data set. The raw data set is scaled by 6: 2: and 2, dividing to respectively obtain a training set, a verification set and a test set. Training set data is used for training the neural network, verifying set data is used for optimizing the neural network, and testing set data evaluates the finally generated network model.
The specific process of preprocessing the plurality of pieces of data of the database in the first step is as follows: and the fixed label of one piece of data is kept unchanged, the fixed characteristic parameter of the data is also kept unchanged, and the period characteristic parameters which do not exceed the time node are accumulated according to the time node of the life cycle to obtain the accumulated period characteristic parameters. And then, carrying out noise value elimination and data dimensionless processing on the data containing the fixed labels, the fixed characteristic parameters and the accumulated period characteristic parameters to obtain an original data set.
Step three: and establishing a road use performance prediction neural network model.
Constructing an original artificial neural network model structure, wherein the original artificial neural network model structure comprises an input layer, a hidden layer and an output layer, the number of neurons of the input layer corresponds to the sum of the index item number of the fixed characteristic parameter and the index item number of the operation period data in the accumulated period characteristic parameter, and the number of neurons of the output layer corresponds to the index item number of the use performance data in the accumulated period characteristic parameter; the number of layers of the hidden layer and the number of the neurons in each hidden layer are randomly selected in respective value ranges, the value range of the number of the hidden layers is 2-50, and the value range of the number of the neurons in each hidden layer is 2-1000. And simultaneously setting an activation function of each layer network layer and a loss function of the model, wherein the input layer and the output layer adopt 'linear' as the activation function, each hidden layer adopts 'relu' as the activation function, and the loss function of the model is set as a 'mae' function. The overfitting phenomenon is avoided, meanwhile, regularization processing is carried out on the hidden layer, regularization is achieved by adding a regularization option and a discarding layer in the hidden layer, at least one of the regularization option is selected randomly in weight regularization, bias regularization and output regularization, and the discarding probability of the discarding layer is selected randomly in the range of 0-0.5. The bias adopts zero initialization, namely the bias among the neurons is set to be 0; the weights are initialized by adopting standard normal distribution, namely, the weight values are independent from each other and obey the distribution with the mean value of 0 and the variance of 1.
Inputting the index of the fixed characteristic parameter in the training set data and the characteristic value of the index of the operation period data in the accumulated period characteristic parameter in the step two into the constructed original artificial neural network model to obtain the predicted data of the index of the use performance data in the accumulated period characteristic parameter calculated by the original artificial neural network model, comparing the calculated predicted data with the characteristic value of the index of the use performance data in the accumulated period characteristic parameter in the training set to obtain an error signal, reversely adjusting the weight and the bias between the input layer and the hidden layer, between the hidden layer and the output layer according to the error signal, repeatedly iterating, and continuously adjusting the weight and the bias to reduce the value of the loss function until the average loss does not decrease any more in the process of iterating and calculating one hundred times by the neural network model under the weight and the bias, namely, the training is stopped to obtain the trained neural network model.
Evaluating the trained neural network model: inputting the characteristic values of the indexes of the fixed characteristic parameters and the indexes of the operation period data in the accumulated period characteristic parameters in the verification set into the trained neural network model to obtain the predicted values of the use performance data of the verification set; then carrying out anti-dimensionless treatment on the predicted value of the service performance data of the verification set, calculating an evaluation index value of the model according to the predicted value and the actual value of the service performance data, and selecting an MAPE value (mean absolute percentage error) as an evaluation index of the model:
Figure BDA0003082906030000041
in the formula: n is the number of data pieces of the verification set, and i is the ith data piece; y isiA true value for an index of the performance data for the validation set;
Figure BDA0003082906030000042
a predicted value of an index is used for the performance data of the verification set. Calculating to obtain the MAPE value of one index of the service performance data of the verification set according to the formula, wherein the MAPE value of the verification set under the model is the average MAPE value of multiple indexes and is set as the MAPE reference value;
an optimization threshold is set, typically in the range of 0-0.4, i.e., the magnitude of the drop in the expected mean MAPE value. And optimizing the trained neural network model by using the verification set data in combination with the MAPE reference value, wherein the specific method comprises the following steps: setting the value range and the value interval of each hyper-parameter, taking the number of hidden layer layers, the number of neurons of each hidden layer, the number of neuron discarding layer layers, regularization options, batch processing data volume and the neuron discarding layer discarding proportion as the hyper-parameters to be optimized, and respectively obtaining various values of each hyper-parameter according to the value range and the value interval of each hyper-parameter; combining multiple values of different hyper-parameters randomly and repeatedly to obtain multiple sets of hyper-parameter data; substituting a group of hyper-parameter data into a trained neural network model, initializing bias by adopting zero, initializing weight by adopting standard normal distribution, inputting the indexes of fixed characteristic parameters in verification set data and the characteristic values of the indexes of operation period data in accumulated period characteristic parameters into an input layer, and obtaining the prediction data of the indexes of use performance data in the accumulated period characteristic parameters through calculation of the neural network model; and comparing the calculated prediction data with the actual values in the verification set to obtain an error signal. And reversely adjusting the weight and the bias between the input layer and the hidden layer, between the hidden layer and between the hidden layer and the output layer according to the error signal, repeatedly iterating, and continuously adjusting the weight and the bias to reduce the value of the loss function until the average loss does not decrease in the iteration process for one hundred times, and terminating the iteration to obtain the neural network model under the current super parameter combination. Calculating the MAPE value of the neural network model, calculating the descending amplitude by combining with the MAPE reference value, continuously selecting the next group of hyper-parameter combination to train the neural network model if the descending amplitude is smaller than the optimization threshold, and continuously repeating the process until the descending amplitude of the MAPE value of the generated neural network model is larger than the optimization threshold, wherein the neural network model under the group of hyper-parameter data is the road use performance prediction neural network model.
The characteristic values of the indexes of the fixed characteristic parameters of the test set data and the indexes of the operation period data in the accumulated period characteristic parameters are put into a road usability prediction neural network model, the precision of the model is evaluated according to the usability data obtained by prediction, the evaluation indexes are an RMSE value (root mean square error) and an MAPE value, and the calculation formula of the RMSE value is as follows:
Figure BDA0003082906030000051
in the formula: n is the number of data pieces of the test set, and i is the ith data piece; y isiA true value for an index of performance data for the test set;
Figure BDA0003082906030000061
a predicted value of an index is used as the performance data of the test set. And calculating to obtain the RMSE value of one index of the service performance data of the test set according to the formula. And when the MAPE values of all indexes are less than 0.07, and the RMSE values of all indexes are not more than ten percent of the full value of the corresponding index item, the model is an effective model. The full value of an index is the maximum value of the true value of the corresponding item in the test set.
Step four: providing maintenance decision for target operation highway by using road use performance prediction neural network model obtained in step three
And (3) acquiring the traffic condition indexes of the operation period data and the characteristic values of the environmental factor indexes of the operation period data in the fixed characteristic parameters and the periodic characteristic parameters of the target operation highway according to the data acquisition method in the step one, and preprocessing the acquired characteristic values according to the data preprocessing process in the step two to form historical data. And inputting the historical data into the road use performance prediction neural network model obtained in the third step, and calculating to obtain the prediction data of the use performance data in the accumulated period characteristic parameters of the road surface of the corresponding hectometer stake mark interval of the target highway. And providing data support for maintenance decision according to the calculated prediction data.
The concrete maintenance decision process is as follows: summarizing the predicted data of the service performance data in the accumulated period characteristic parameters of all hectometer pile numbers of the target time nodes (target years), giving out performance weight and grading basis according to the conditions of local expressways, weighting all service performance to obtain the comprehensive performance indexes of all hectometer pile numbers, grading the comprehensive performance of different pile numbers according to the grading basis, selecting the maintenance standard grade, wherein the pile numbers under the grade are the maintenance pile numbers, and making decisions according to the refined service performance by specific maintenance measures.
Step five: planning the construction period data of the highway in the construction period by using the road use performance prediction model obtained in the step three and combining a genetic algorithm, wherein the specific process is as follows:
(1) and (3) according to the traffic planning of the highway in the construction period, the local historical environmental factors and the service life planning, carrying out data arrangement by adopting the method in the first step and preprocessing by adopting the method in the second step to obtain an accumulated period characteristic parameter data sequence in the planned life period a. The traffic condition index of the operation period data is determined according to the traffic plan and the service life plan, and the environmental factor index of the operation period data is predicted according to local historical environmental factors (the prediction method refers to the prior art, and a common exponential smoothing method is adopted).
(2) Setting parameters in a genetic algorithm, specifically variable parameters, fixed parameters and characterization parameters, setting material attributes as variable parameters, setting mechanical indexes as fixed parameters, setting accumulated cycle characteristic parameter data in a planning life cycle a as characterization parameters, and obtaining characterization parameter data through the step (1). According to construction planning of a highway in a construction period, obtaining a mechanical index of data in the construction period and a reference value range of a material attribute index of the data in the construction period, namely obtaining a value range of each index of a variable parameter and a value range of each index of a fixed parameter, uniformly selecting 20 values of each index of the variable parameter in the respective value range, and then randomly combining the values of the indexes of the variable parameter until 200 different variable parameter combinations are generated. And randomly selecting one value from each index of the fixed parameters in the respective value range to form a fixed parameter combination.
Combining a group of variable parameter combinations, fixed parameter combinations and characterization parameters, and using the combined values as a group of inputs of a road use performance prediction neural network model in the third step to obtain a predicted value of use performance data; the fitness function is set to:
Figure BDA0003082906030000071
Figure BDA0003082906030000072
wherein: q is the fitness of an individual, where an individual refers to a set of inputs; a is the number of time periods; y istThe comprehensive fitness output for the t time node of the prediction model; n is the number of output layers of the prediction model, namely the number of neurons in the output layer; y isiThe output of the output layer neuron i of the prediction model corresponds to the predicted value of the performance index i. The lambda is the important coefficient of the use performance, the value range of the lambda is between 0 and 1, and the sum of all the important coefficients of the performance is 1.
And respectively putting a plurality of groups of inputs consisting of different groups of variable parameter combinations, fixed parameter combinations and characterization parameters into the prediction model to obtain predicted values of the service performance data corresponding to the different groups of inputs. Calculating the fitness of each group of input, and setting the maximum fitness of individuals in the whole population as Q1。
(3) The variable parameters are encoded in a floating-point number encoding mode, and each index of the variable parameters occupies one chromosome position and is stored in a floating-point number type. Taking 200 different groups of variable parameter combinations in the step (2) as a first generation population, sorting the fitness of individuals formed by the first generation population from large to small, removing fifty percent of the individuals, and then selecting the first 20 individuals with the ratio value sorted from large to small as operators according to a fitness ratio method for the rest individuals, namely selecting the probability that the fitness of the individual accounts for the fitness of all the rest individuals:
Figure BDA0003082906030000081
wherein: p is the probability of selecting the individual as an operator; and Q is the fitness.
Carrying out two-group random matching on the selected 20 operators to obtain 10 operator pairs; each set of operator pairs is interleaved in a consistent hybridization fashion (i.e., genes at each locus of two paired individuals are swapped with the same crossover probability to form new individuals) until 150 new individuals are generated. Performing mutation operation on 150 new individuals, firstly setting the mutation probability to be 0.05, selecting the individuals to be mutated, randomly selecting mutation sites for performing real value mutation on the individuals to be mutated, and replacing the new individuals subjected to the mutation operation with the individuals to be mutated to obtain 150 offspring individuals; the 150 offspring individuals plus the first 25% of the individuals in the first generation population are added to generate a second generation population. And the third generation population is obtained on the basis of the second generation population according to the generation process of the second generation population. The objects of the cross operation and the mutation operation are variable parameters in a group of input, the fixed parameters and the characterization parameters are fixed values, and the fixed parameters and the characterization parameters only relate to the calculation of individual fitness.
(4) After each new generation of population is generated, determining whether to continuously update the population by iteration or terminate iteration according to an iteration termination condition; the termination conditions are specifically:
1) the preset termination fitness is set as QmSetting the maximum fitness of the nth generation population as QnN is not less than 2, and Q is judgednWhether or not it is greater than QmIf yes, executing step 2), otherwise, executing step 3);
2) increase of preset fitnessAmplitude threshold is eta, judge QnComparison of Qn-1If the amplification is less than eta, if so, the iteration is terminated, and Q and the output are outputnCorresponding individuals, otherwise, executing the step 3);
3) the threshold value of the maximum algebra of the preset population is Gm(i.e., the threshold number of iterative updates is Gm-1), judging whether the current population algebra reaches the maximum algebra, if so, terminating iteration and outputting the maximum fitness Q in the current populationnOtherwise, generating an n +1 generation population according to the step (4) on the basis of the n generation population, skipping to the step 1), and enabling the maximum fitness Q of the n +1 generation populationn+1And QmPerforming size judgment to determine whether to execute the step 2) or the step 3) next;
(5) decoding the individual corresponding to the maximum fitness in the population meeting the iteration termination condition in the step (4) to obtain the optimal value of each index of the variable parameter, namely the optimal value of each index of the material attribute;
(6) setting the material attribute as a fixed parameter, wherein the value of the fixed parameter combination is the optimal value of each index of the material attribute obtained in the step (5); and (5) setting the mechanical index as a variable parameter, keeping the representation parameter unchanged, and repeating the steps (2) to (5) to obtain the optimal value of the mechanical index.
Compared with the prior art, the invention has the beneficial effects that:
1. the method comprehensively examines various indexes of the whole life cycle of the road, respectively processes data according to different index attributes, establishes a database which has wide adaptability and can comprehensively reflect road characteristics, and lays a foundation for maintenance decision or construction guidance of the road.
2. And establishing a service performance prediction model of the road by using the database and the artificial neural network, and providing data reference for maintenance decision of the road.
3. Providing a genetic algorithm of semi-parameter search, performing global search on the characteristic value combination by using a road use performance prediction model, performing parameter optimization on the mechanical index and the material attribute of the highway in the construction period, and finely controlling the construction parameters; and the applied three-layer termination control method can excavate the optimal variable parameters to the maximum extent, and meanwhile, the optimal variable parameters cannot be trapped in an infinite loop situation, and finally the search of the optimal construction parameters is completed.
4. The tracing method inspects various indexes of the whole life cycle of the road, comprehensively and systematically utilizes and analyzes data, and links the accurate maintenance and scientific construction of the expressway together, so that a prediction model and feedback analysis have integrity, the construction cost and the maintenance cost of the life cycle of the expressway are reduced from a macroscopic view, and the construction efficiency and the operation efficiency are improved.
Drawings
Fig. 1 is a diagram of a road use performance prediction neural network model structure according to an embodiment of the tracing method of the present invention.
Fig. 2 is a flowchart of planning construction period data of a highway in a construction period by using a road use performance prediction model in combination with a genetic algorithm according to an embodiment of the tracing method.
Fig. 3 is a schematic flow chart of an embodiment of the tracing method of the present invention.
Detailed Description
Specific examples of the present invention are given below. The specific examples are only intended to illustrate the invention in further detail and do not limit the scope of protection of the claims of the present application.
The invention constructs a highway pavement full life cycle quality tracing method (a tracing method for short) based on artificial intelligence, which can provide powerful guidance for maintenance of an operation highway and construction of raising the highway, and the tracing method comprises the following steps:
the method comprises the following steps: and collecting the full life cycle data of the highway and establishing a database.
The highway full life cycle data specifically comprises construction period data, operation period data and use performance data, the construction period data specifically comprises mechanical indexes and material attribute indexes, and the operation period data specifically comprises traffic condition indexes and environmental factor indexes.
One piece of data in the database takes a name of a highway, a pile number of a starting point, a road grade, planned traffic volume and a region as a fixed label, takes full life cycle data of a road surface between one hundred meters of pile numbers on one lane as a characteristic parameter, one value of the characteristic parameter is a characteristic value of one index in the full life cycle data of the road surface in one time period, the characteristic value is a data point, and the characteristic value of one index in one time period is obtained according to the data in the corresponding time period;
specifically, the mechanical indexes of the construction period data include: 1. the roadbed deflection, namely the rebound deflection at different positions is measured by adopting a Beckman beam method; 2. the compactness of each layer specifically comprises roadbed compactness, water stability lamination compactness and pavement compactness; 3. the cubic compressive strength of cement mortar specifically comprises the average values of the breaking load and the compressive strength of the cubic body made of the cement mortar at different mixing ratios of each key part.
The material property indexes of the construction period data comprise: 1. and (3) screening results of the soil sample, wherein the results comprise coarse screening results and fine screening results, the coarse screening results are the percentage of the total soil mass under each large aperture, and the large apertures are respectively 2mm, 5mm and 10 mm. The result of the fine screening is the percentage of the total soil mass under each small aperture, and the small apertures are 1mm, 0.5mm, 0.25mm and 0.075mm respectively; 2. the glue-sand ratio of the water stable layer; 3. the California Bearing Ratio (CBR) of the roadbed is the ratio of the unit pressure when the penetration amount of the roadbed soil reaches 2.5mm to the standard load when the standard crushed stone is pressed into the roadbed with the same penetration amount.
The traffic condition indexes of the operation period data include: 1. the traffic axle load is that the traffic volume is converted into equivalent traffic axle load according to the type of traffic unit; 2. the passenger-cargo ratio can assist the equivalent axle load to distinguish the influence of different types of vehicles on the road surface damage.
The environmental factor indexes of the operation period data comprise: 1. hydrological factors mainly including the influence of precipitation and humidity environment; 2. atmospheric factors including atmospheric pressure, wind speed and temperature in the particular area.
The indices of the service performance data include road surface damage condition (PCI), road surface Running Quality (RQI), road surface Rutting (RDI), road surface Skid Resistance (SRI), road surface structural strength (PSSI), road surface service Performance (PQI), and road surface breakage reduced area. The performance data indexes are detected and recorded according to the road technical condition evaluation standard (JTG H20-2007).
The characteristic parameters comprise fixed characteristic parameters and periodic characteristic parameters, the indexes of the construction period data are the fixed characteristic parameters, and the characteristic values are recorded only once in the whole life cycle of the road surface; the cycle characteristic parameters comprise traffic condition indexes of operation period data, environment factor indexes of the operation period data and indexes of using performance data, and the cycle characteristic parameters are respectively recorded in different time periods, namely the recording times of the cycle characteristic parameters are the number of time periods included in the full life cycle.
The characteristic value of the traffic condition index of the operation period data is obtained according to actual data in the corresponding time period, and the characteristic value of the environmental factor index of the operation period data is a comprehensive value of the detection data in the corresponding time period. And obtaining the characteristic value of the index using the performance data according to the detection result in the corresponding time period. The characteristic values of an index in different time periods are obtained in the same way. The characteristic values respectively obtained in different time periods are period characteristic parameters, and the period characteristic parameters are respectively recorded in different time periods, namely the recording times of the period characteristic parameters are the number of time periods included in the full life cycle.
For example, the temperature index, and the time period is one year, the characteristic value of the temperature index in one time period can be calculated as follows:
Figure BDA0003082906030000111
Figure BDA0003082906030000112
in the formula: t is tmIs the monthly temperature of the road;
Figure BDA0003082906030000113
the lowest and highest monthly average temperatures for the road;
Figure BDA0003082906030000114
the lowest and highest lunar extreme temperatures for the road; t is tyIs the annual temperature of the road; α and β are weighting coefficients, which are 0.2 and 0.3, respectively. If the whole life cycle is 10 years, calculating according to the method, and recording the characteristic value of the temperature index of each year.
The characteristic value of one index is corresponding data of a road surface between one hundred-meter pile numbers, partial data is obtained in a road surface interval smaller than one hundred meters, if roadbed deflection and roadbed compaction degree in a construction period exist, weighting processing is carried out on the data in the interval length, weight is distributed on data numerical values according to the length of the interval, and finally summation calculation is carried out to obtain the road surface data of the hundred-meter pile numbers. And acquiring partial data of a road surface interval larger than one hundred meters, such as a soil sample screening result, or acquiring data of a road surface interval smaller than one hundred meters, namely spanning two hundred-meter pile number intervals, corresponding the data larger than one hundred-meter pile number with other data in a repeated value mode, and performing weighted segmentation on the spanned partial data according to the length of the spanned interval.
Step two: and (4) preprocessing a plurality of pieces of data of the database in the step one to obtain an original data set. The raw data set is scaled by 6: 2: and 2, dividing to respectively obtain a training set, a verification set and a test set. Training set data is used for training the neural network, verifying set data is used for optimizing the neural network, and testing set data evaluates the finally generated network model.
The specific process of preprocessing the plurality of pieces of data of the database in the first step is as follows: and the fixed label of one piece of data is kept unchanged, the fixed characteristic parameter of the data is also kept unchanged, and the period characteristic parameters which do not exceed the time node are accumulated according to the time node of the life cycle to obtain the accumulated period characteristic parameters. For example, the accumulated characteristic parameter corresponding to the second time node of the life cycle is obtained by summing the characteristic values of the corresponding index items of the characteristic parameter of the cycle in the first time period and the characteristic parameter of the cycle in the second time period, and the accumulated characteristic parameter corresponding to the last time node is obtained by summing all the characteristic parameters of the cycle in all the time periods. If the time period is one year, the fifth year traffic axle load value is the accumulated value of the previous four years plus the original fifth year traffic axle load. And then, carrying out noise value elimination and data dimensionless processing on the data containing the fixed labels, the fixed characteristic parameters and the accumulated period characteristic parameters to obtain an original data set. (for the accumulated period characteristic parameter, the data corresponding to one index is a data sequence arranged according to the life period node sequence).
The noise value elimination is to remove abnormal values, and the realization method is to remove extreme values of the boxcar graph. The non-dimensionalization comprises normalization and standardization, and an appropriate non-dimensionalization method is selected to process the data.
Step three: and establishing a road use performance prediction neural network model.
Constructing an original artificial neural network model structure, wherein the original artificial neural network model structure comprises an input layer, a hidden layer and an output layer, the number of neurons of the input layer corresponds to the sum of the index item number of the fixed characteristic parameter and the index item number of the operation period data in the accumulated period characteristic parameter, and the number of neurons of the output layer corresponds to the index item number of the use performance data in the accumulated period characteristic parameter; the number of layers of the hidden layer and the number of the neurons in each hidden layer are randomly selected in respective value ranges, the value range of the number of the hidden layers is 2-50, and the value range of the number of the neurons in each hidden layer is 2-1000. And simultaneously setting an activation function of each layer network layer and a loss function of the model, wherein the input layer and the output layer adopt 'linear' as the activation function, each hidden layer adopts 'relu' as the activation function, and the loss function of the model is set as a 'mae' function. The overfitting phenomenon is avoided, meanwhile, regularization processing is carried out on the hidden layer, regularization is achieved by adding a regularization option and a discarding layer in the hidden layer, at least one of the regularization option is selected randomly in weight regularization, bias regularization and output regularization, and the discarding probability of the discarding layer is selected randomly in the range of 0-0.5. The bias adopts zero initialization, namely the bias among the neurons is set to be 0; the weights are initialized by adopting standard normal distribution, namely, the weight values are independent from each other and obey the distribution with the mean value of 0 and the variance of 1.
Inputting the index of the fixed characteristic parameter in the training set data and the characteristic value of the index of the operation period data in the accumulated period characteristic parameter in the step two into the constructed original artificial neural network model to obtain the predicted data of the index of the use performance data in the accumulated period characteristic parameter calculated by the original artificial neural network model, comparing the calculated predicted data with the characteristic value of the index of the use performance data in the accumulated period characteristic parameter in the training set to obtain an error signal, reversely adjusting the weight and the bias between the input layer and the hidden layer, between the hidden layer and the output layer according to the error signal, repeatedly iterating, and continuously adjusting the weight and the bias to reduce the value of the loss function until the average loss does not decrease any more in the process of iterating and calculating one hundred times by the neural network model under the weight and the bias, namely, the training is stopped to obtain the trained neural network model.
Evaluating the trained neural network model: inputting the characteristic values of the indexes of the fixed characteristic parameters and the indexes of the operation period data in the accumulated period characteristic parameters in the verification set into the trained neural network model to obtain the predicted values of the use performance data of the verification set; then carrying out anti-dimensionless treatment on the predicted value of the service performance data of the verification set, calculating an evaluation index value of the model according to the predicted value and the actual value of the service performance data, and selecting an MAPE value (mean absolute percentage error) as an evaluation index of the model:
Figure BDA0003082906030000141
in the formula: n is the number of data pieces of the verification set, and i is the ith data piece; y isiA true value for an index of the performance data for the validation set;
Figure BDA0003082906030000142
a predicted value of an index is used for the performance data of the verification set. Calculating to obtain MAPE value of one index of service performance data of the verification set according to the formula, wherein the MAPE value of the verification set under the model is of multiple indexesAveraging the MAPE value and setting the MAPE value as a MAPE reference value;
an optimization threshold is set, typically in the range of 0-0.4, i.e., the magnitude of the drop in the expected mean MAPE value. And optimizing the trained neural network model by using the verification set data in combination with the MAPE reference value, wherein the specific method comprises the following steps: setting the value range and the value interval of each hyper-parameter, taking the number of hidden layer layers, the number of neurons of each hidden layer, the number of neuron discarding layer layers, regularization options, batch processing data volume and the neuron discarding layer discarding proportion as the hyper-parameters to be optimized, and respectively obtaining various values of each hyper-parameter according to the value range and the value interval of each hyper-parameter; combining multiple values of different hyper-parameters randomly and repeatedly to obtain multiple sets of hyper-parameter data; substituting a group of hyper-parameter data into a trained neural network model, initializing bias by adopting zero, initializing weight by adopting standard normal distribution, inputting the indexes of fixed characteristic parameters in verification set data and the characteristic values of the indexes of operation period data in accumulated period characteristic parameters into an input layer, and obtaining the prediction data of the indexes of use performance data in the accumulated period characteristic parameters through calculation of the neural network model; and comparing the calculated prediction data with the actual values in the verification set to obtain an error signal. And reversely adjusting the weight and the bias between the input layer and the hidden layer, between the hidden layer and between the hidden layer and the output layer according to the error signal, repeatedly iterating, and continuously adjusting the weight and the bias to reduce the value of the loss function until the average loss does not decrease in the iteration process for one hundred times, and terminating the iteration to obtain the neural network model under the current super parameter combination. Calculating the MAPE value of the neural network model, calculating the descending amplitude by combining with the MAPE reference value, continuously selecting the next group of hyper-parameter combination to train the neural network model if the descending amplitude is smaller than the optimization threshold, and continuously repeating the process until the descending amplitude of the MAPE value of the generated neural network model is larger than the optimization threshold, wherein the neural network model under the group of hyper-parameter data is the road use performance prediction neural network model.
The characteristic values of the indexes of the fixed characteristic parameters of the test set data and the indexes of the operation period data in the accumulated period characteristic parameters are put into a road use performance prediction neural network model, the precision of the model is evaluated according to the predicted use performance data, the evaluation indexes are an RMSE value and an MAPE value, and the calculation formula of the RMSE value is as follows:
Figure BDA0003082906030000151
in the formula: n is the number of data pieces of the test set, and i is the ith data piece; y isiA true value for an index of performance data for the test set;
Figure BDA0003082906030000152
a predicted value of an index is used as the performance data of the test set. And calculating to obtain the RMSE value of one index of the service performance data of the test set according to the formula. And when the MAPE values of all indexes are less than 0.07, and the RMSE values of all indexes are not more than ten percent of the full value of the corresponding index item, the model is an effective model. The full value of an index is the maximum value of the true value of the corresponding item in the test set.
Step four: providing maintenance decision for target operation highway by using road use performance prediction neural network model obtained in step three
And (3) acquiring the traffic condition indexes of the operation period data and the characteristic values of the environmental factor indexes of the operation period data in the fixed characteristic parameters and the periodic characteristic parameters of the target operation highway according to the data acquisition method in the step one, and preprocessing the acquired characteristic values according to the data preprocessing process in the step two to form historical data. And inputting the historical data into the road use performance prediction neural network model obtained in the third step, and calculating to obtain prediction data of indexes of use performance data in the accumulated period characteristic parameters of the road surface of the corresponding hectometer stake mark interval of the target highway. And providing data support for maintenance decision according to the calculated prediction data.
The concrete process of maintenance decision is as follows: summarizing the predicted data of the indexes of the service performance data in the accumulated period characteristic parameters of all hectometer pile numbers of the target time nodes (target years), giving out performance weight and grading basis according to the conditions of local expressways, weighting all service performance to obtain the comprehensive performance indexes of all hectometer pile numbers, grading the comprehensive performance of different pile numbers according to the grading basis, selecting the maintenance standard grade, wherein the pile numbers under the grade are the maintenance pile numbers, and the specific maintenance measures are decided according to the refined service performance.
Step five: planning the construction period data of the highway in the construction period by using the road use performance prediction model obtained in the step three and combining a genetic algorithm, wherein the specific process is as follows:
(1) and (3) according to the traffic planning of the highway in the construction period, the local historical environmental factors and the service life planning, carrying out data arrangement by adopting the method in the first step and preprocessing by adopting the method in the second step to obtain an accumulated period characteristic parameter data sequence in the planned life period a. The traffic condition index of the operation period data is determined according to the traffic plan and the service life plan, and the environmental factor index of the operation period data is predicted according to local historical environmental factors (the prediction method refers to the prior art, and a common exponential smoothing method is adopted).
(2) Setting parameters in a genetic algorithm, specifically variable parameters, fixed parameters and characterization parameters, setting material attributes as variable parameters, setting mechanical indexes as fixed parameters, setting accumulated cycle characteristic parameter data in a planning life cycle a as characterization parameters, and obtaining characterization parameter data through the step (1). According to construction planning of a highway in a construction period, obtaining a mechanical index of data in the construction period and a reference value range of a material attribute index of the data in the construction period, namely obtaining a value range of each index of a variable parameter and a value range of each index of a fixed parameter, uniformly selecting 20 values of each index of the variable parameter in the respective value range, and then randomly combining the values of the indexes of the variable parameter until 200 different variable parameter combinations are generated. And randomly selecting one value from each index of the fixed parameters in the respective value range to form a fixed parameter combination.
Combining a group of variable parameter combinations, fixed parameter combinations and characterization parameters, and using the combined values as a group of inputs of a road use performance prediction neural network model (prediction model for short) in the third step to obtain a predicted value of a target value; the fitness function is set to:
Figure BDA0003082906030000161
Figure BDA0003082906030000162
wherein: q is the fitness of an individual, where an individual refers to a set of inputs; a is the number of time periods; y istThe comprehensive fitness output for the t time node of the prediction model; n is the number of output layers of the prediction model, namely the number of neurons in the output layer; y isiThe output of the output layer neuron i of the prediction model corresponds to the predicted value of the performance index i. The lambda is the important coefficient of the use performance, the value range of the lambda is between 0 and 1, and the sum of all the important coefficients of the performance is 1.
And respectively putting a plurality of groups of inputs consisting of different groups of variable parameter combinations, fixed parameter combinations and characterization parameters into the prediction model to obtain predicted values of the target values corresponding to the different groups of inputs. Calculating the fitness of each group of input, and setting the maximum fitness of individuals in the whole population as Q1。
(3) The variable parameters are encoded in a floating-point number encoding mode, and each index of the variable parameters occupies one chromosome position and is stored in a floating-point number type. Taking 200 different groups of variable parameter combinations in the step (2) as a first generation population, sorting the fitness of individuals formed by the first generation population from large to small, removing fifty percent of the individuals, and then selecting the first 20 individuals with the ratio value sorted from large to small as operators according to a fitness ratio method for the rest individuals, namely selecting the probability that the fitness of the individual accounts for the fitness of all the rest individuals:
Figure BDA0003082906030000171
wherein: p is the probability of selecting the individual as an operator; and Q is the fitness.
Carrying out two-group random matching on the selected 20 operators to obtain 10 operator pairs; each set of operator pairs is interleaved in a consistent hybridization fashion (i.e., genes at each locus of two paired individuals are swapped with the same crossover probability to form new individuals) until 150 new individuals are generated. Performing mutation operation on 150 new individuals, firstly setting the mutation probability to be 0.05, selecting the individuals to be mutated, randomly selecting mutation sites for performing real value mutation on the individuals to be mutated, and replacing the new individuals subjected to the mutation operation with the individuals to be mutated to obtain 150 offspring individuals; the 150 offspring individuals plus the first 25% of the individuals in the first generation population are added to generate a second generation population. And the third generation population is obtained on the basis of the second generation population according to the generation process of the second generation population. The objects of the cross operation and the mutation operation are variable parameters in a group of input, the fixed parameters and the characterization parameters are fixed values, and the fixed parameters and the characterization parameters only relate to the calculation of individual fitness.
(4) After each new generation of population is generated, determining whether to continuously update the population by iteration or terminate iteration according to an iteration termination condition; the termination conditions are specifically:
1) the preset termination fitness is set as QmSetting the maximum fitness of the nth generation population as QnN is not less than 2, and Q is judgednWhether or not it is greater than QmIf yes, executing step 2), otherwise, executing step 3);
2) presetting a fitness amplification threshold value as eta, and judging QnComparison of Qn-1If the amplification is less than eta, if so, the iteration is terminated, and Q and the output are outputnCorresponding individuals, otherwise, executing the step 3);
3) the threshold value of the maximum algebra of the preset population is Gm(i.e., the threshold number of iterative updates is Gm-1), judging whether the current population algebra reaches the maximum algebra, if so, stopping iteration and outputting the maximum algebra in the current populationFitness QnOtherwise, generating an n +1 generation population according to the step (4) on the basis of the n generation population, skipping to the step 1), and enabling the maximum fitness Q of the n +1 generation populationn+1And QmPerforming size judgment to determine whether to execute the step 2) or the step 3) next;
(5) decoding the individual corresponding to the maximum fitness in the population meeting the iteration termination in the step (4) to obtain the optimal value of each index of the variable parameter, namely the optimal value of each index of the material attribute;
(6) setting the material attribute as a fixed parameter, wherein the value of the fixed parameter combination is the optimal value of each index of the material attribute obtained in the step (5); and (5) setting the mechanical index as a variable parameter, keeping the representation parameter unchanged, and repeating the steps (2) to (5) to obtain the optimal value of the mechanical index.
The invention relates to a method for tracing the quality of a road surface in a full life cycle based on artificial intelligence. On the basis of the performance prediction model, maintenance decisions are provided for the expressway in the operation period, and construction period data of the expressway in the construction period can be planned by combining a genetic algorithm to guide construction. The invention takes various index data of the whole life cycle of the road as an investigation object, respectively processes the data according to different attributes of each index, integrates the data into a database, and realizes the whole life cycle tracing of the highway by combining an artificial neural network and a genetic algorithm.
Nothing in this specification is said to apply to the prior art.

Claims (7)

1. A full life cycle quality tracing method for a road pavement based on artificial intelligence is characterized by comprising the following steps:
the method comprises the following steps: collecting full life cycle data of the highway and establishing a database;
the highway full life cycle data specifically comprises construction period data, operation period data and use performance data, the construction period data specifically comprises mechanical indexes and material attribute indexes, and the operation period data specifically comprises traffic condition indexes and environmental factor indexes;
one piece of data in the database takes a name of a highway, a pile number of a starting point, a road grade, planned traffic volume and a region as a fixed label, takes full life cycle data of a road surface between one hundred meters of pile numbers on one lane as a characteristic parameter, one value of the characteristic parameter is a characteristic value of one index in the full life cycle data of the road surface in one time period, the characteristic value is a data point, and the characteristic value of one index in one time period is obtained according to the data in the corresponding time period;
the characteristic parameters comprise fixed characteristic parameters and periodic characteristic parameters, the indexes of the construction period data are the fixed characteristic parameters, and the characteristic values are recorded only once in the whole life cycle of the road surface; the cycle characteristic parameters comprise traffic condition indexes of operation period data, environment factor indexes of the operation period data and indexes of using performance data, and are respectively recorded in different time periods, namely the recording times of the cycle characteristic parameters are the number of time periods included in the full life cycle;
step two: preprocessing a plurality of pieces of data of the database in the step one to obtain an original data set; the raw data set is scaled by 6: 2: 2, dividing to respectively obtain a training set, a verification set and a test set; training set data is used for training a neural network, verifying set data is used for optimizing the neural network, and testing set data evaluates a finally generated network model;
the specific process of preprocessing the plurality of pieces of data of the database in the first step is as follows: the fixed label of one piece of data is kept unchanged, the fixed characteristic parameter of the data is also kept unchanged, and the period characteristic parameters which do not exceed the time node are accumulated according to the time node of the life cycle to obtain accumulated period characteristic parameters; then, carrying out noise value elimination and data dimensionless processing on the data containing the fixed label, the fixed characteristic parameter and the accumulated period characteristic parameter to obtain an original data set;
step three: establishing a road use performance prediction neural network model;
constructing an original artificial neural network model structure, wherein the original artificial neural network model structure comprises an input layer, a hidden layer and an output layer, the number of neurons of the input layer corresponds to the sum of the index item number of the fixed characteristic parameter and the index item number of the operation period data in the accumulated period characteristic parameter, and the number of neurons of the output layer corresponds to the index item number of the use performance data in the accumulated period characteristic parameter; the number of layers of the hidden layer and the number of neurons in each hidden layer are randomly selected in respective value ranges, the value range of the number of the hidden layers is 2-50, and the value range of the number of the neurons in each hidden layer is 2-1000; simultaneously setting an activation function of each layer network layer and a loss function of the model, wherein the input layer and the output layer adopt 'linear' as the activation function, each hidden layer adopts 'relu' as the activation function, and the loss function of the model is set as a 'mae' function; the overfitting phenomenon is avoided, meanwhile, regularization processing is carried out on the hidden layer, regularization is achieved by adding a regularization option and a discarding layer in the hidden layer, at least one item of the regularization option is randomly selected in weight regularization, bias regularization and output regularization, and the discarding probability of the discarding layer is randomly selected in the range of 0-0.5; the bias adopts zero initialization, namely the bias among the neurons is set to be 0; the weights are initialized by adopting standard normal distribution, namely, the weights are independent from each other and obey the distribution with the mean value of 0 and the variance of 1;
inputting the index of the fixed characteristic parameter in the training set data and the characteristic value of the index of the operation period data in the accumulated period characteristic parameter in the step two into the constructed original artificial neural network model to obtain the predicted data of the index of the use performance data in the accumulated period characteristic parameter calculated by the original artificial neural network model, comparing the calculated predicted data with the characteristic value of the index of the use performance data in the accumulated period characteristic parameter in the training set to obtain an error signal, reversely adjusting the weight and the bias between the input layer and the hidden layer, between the hidden layer and the output layer according to the error signal, repeatedly iterating, and continuously adjusting the weight and the bias to reduce the value of the loss function until the average loss does not decrease any more in the process of iterating and calculating one hundred times by the neural network model under the weight and the bias, terminating the training to obtain a trained neural network model;
evaluating the trained neural network model: inputting the characteristic values of the indexes of the fixed characteristic parameters and the indexes of the operation period data in the accumulated period characteristic parameters in the verification set into the trained neural network model to obtain the predicted values of the use performance data of the verification set; then carrying out anti-dimensionless treatment on the predicted value of the service performance data of the verification set, calculating the evaluation index value of the model according to the predicted value and the actual value of the service performance data, and selecting the MAPE value as the evaluation index of the model:
Figure FDA0003082906020000031
in the formula: n is the number of data pieces of the verification set, and i is the ith data piece; y isiA true value for an index of the performance data for the validation set;
Figure FDA0003082906020000032
a predicted value of an index for the performance data of the validation set; calculating to obtain the MAPE value of one index of the service performance data of the verification set according to the formula, wherein the MAPE value of the verification set under the model is the average MAPE value of multiple indexes and is set as the MAPE reference value;
setting an optimization threshold, generally in the range of 0-0.4, which is the expected decrease of the average MAPE value; and optimizing the trained neural network model by using the verification set data in combination with the MAPE reference value, wherein the specific method comprises the following steps: setting the value range and the value interval of each hyper-parameter, taking the number of hidden layer layers, the number of neurons of each hidden layer, the number of neuron discarding layer layers, regularization options, batch processing data volume and the neuron discarding layer discarding proportion as the hyper-parameters to be optimized, and respectively obtaining various values of each hyper-parameter according to the value range and the value interval of each hyper-parameter; combining multiple values of different hyper-parameters randomly and repeatedly to obtain multiple sets of hyper-parameter data; substituting a group of hyper-parameter data into a trained neural network model, initializing bias by adopting zero, initializing weight by adopting standard normal distribution, inputting the indexes of fixed characteristic parameters in verification set data and the characteristic values of the indexes of operation period data in accumulated period characteristic parameters into an input layer, and obtaining the prediction data of the indexes of use performance data in the accumulated period characteristic parameters through calculation of the neural network model; comparing the calculated prediction data with the actual values in the verification set to obtain error signals; reversely adjusting the weights and the offsets between the input layer and the hidden layer, between the hidden layer and the hidden layer, and between the hidden layer and the output layer according to the error signals, repeatedly iterating, and continuously adjusting the weights and the offsets to reduce the value of the loss function, and terminating iteration until the average loss is not reduced in the iteration process for one hundred times to obtain a neural network model under the current super parameter combination; calculating the MAPE value of the neural network model, calculating the descending amplitude by combining with the MAPE reference value, continuously selecting the next group of hyper-parameter combination to train the neural network model if the descending amplitude is smaller than the optimization threshold, and continuously repeating the process until the descending amplitude of the MAPE value of the generated neural network model is larger than the optimization threshold, wherein the neural network model under the group of hyper-parameter data is the road use performance prediction neural network model;
the characteristic values of the indexes of the fixed characteristic parameters of the test set data and the indexes of the operation period data in the accumulated period characteristic parameters are put into a road use performance prediction neural network model, the precision of the model is evaluated according to the predicted use performance data, the evaluation indexes are an RMSE value and an MAPE value, and the calculation formula of the RMSE value is as follows:
Figure FDA0003082906020000041
in the formula: n is the number of data pieces of the test set, and i is the ith data piece; y isiA true value for an index of performance data for the test set;
Figure FDA0003082906020000042
a predicted value of an index of the performance data of the test set; according to the formula, calculating to obtain the RMSE value of one index of the service performance data of the test set; when the MAPE values of all indexes are less than 0.07, and the RMSE values of all indexes are not more than ten percent of the full value of the corresponding index item, the model is an effective model; the full value of one index is the maximum value of the real value of the corresponding item in the test set;
step four: providing a maintenance decision for the target operation highway by using the road use performance prediction neural network model obtained in the step three;
collecting the traffic condition indexes of the operation period data and the characteristic values of the environmental factor indexes of the operation period data in the fixed characteristic parameters and the periodic characteristic parameters of the target operation highway according to the data collection method in the step one, and preprocessing the collected characteristic values according to the data preprocessing process in the step two to form historical data; inputting the historical data into the road use performance prediction neural network model obtained in the third step, and calculating to obtain the prediction data of the use performance data in the accumulated period characteristic parameters of the road surface of the corresponding hectometer stake mark interval of the target highway; providing data support for maintenance decision according to the calculated prediction data;
step five: planning the construction period data of the highway in the construction period by using the road use performance prediction model obtained in the step three and combining a genetic algorithm, wherein the specific process is as follows:
(1) according to the traffic planning of the highway in the construction period, the local historical environmental factors and the service life planning, the method in the first step is adopted for data arrangement, and the method in the second step is adopted for preprocessing, so that an accumulated period characteristic parameter data sequence in the planned life period a is obtained; the traffic condition indexes of the operation period data are determined according to the traffic plan and the service life plan, and the environmental factor indexes of the operation period data are predicted according to local historical environmental factors;
(2) setting parameters in a genetic algorithm, specifically variable parameters, fixed parameters and characterization parameters, setting material attributes as variable parameters, setting mechanical indexes as fixed parameters, setting accumulated period characteristic parameter data in a planning life period a as characterization parameters, and obtaining characterization parameter data through the step (1); according to construction planning of a highway in a construction period, obtaining a mechanical index of data in the construction period and a reference value range of a material attribute index of the data in the construction period, namely obtaining a value range of each index of a variable parameter and a value range of each index of a fixed parameter, uniformly selecting 20 values of each index of the variable parameter in the respective value range, and then randomly combining the values of the indexes of the variable parameter until 200 different variable parameter combinations are generated; randomly selecting a value of each index of the fixed parameters in respective value range to form a fixed parameter combination;
combining a group of variable parameter combinations, fixed parameter combinations and characterization parameters, and using the combined values as a group of inputs of a road use performance prediction neural network model in the third step to obtain a predicted value of use performance data; the fitness function is set to:
Figure FDA0003082906020000051
Figure FDA0003082906020000052
wherein: q is the fitness of an individual, where an individual refers to a set of inputs; a is the number of time periods; y istThe comprehensive fitness output for the t time node of the prediction model; n is the number of output layers of the prediction model, namely the number of neurons in the output layer; y isiOutputting the output layer neuron i of the prediction model corresponding to the predicted value of the service performance index i; lambda is an important coefficient of the use performance, the value range of lambda is between 0 and 1, and the sum of all the important coefficients of the performance is 1;
respectively inputting a plurality of groups of inputs consisting of different groups of variable parameter combinations, fixed parameter combinations and characterization parameters into a prediction model to obtain predicted values of service performance data corresponding to different groups of inputs;calculating the fitness of each group of input, and setting the maximum fitness of individuals in the whole population as Q1；
(3) Encoding the variable parameters by adopting a floating point number encoding mode, wherein each index of the variable parameters occupies a chromosome position and is stored in a floating point number type; taking 200 different groups of variable parameter combinations in the step (2) as a first generation population, sorting the fitness of individuals formed by the first generation population from large to small, removing fifty percent of the individuals, and then selecting the first 20 individuals with the ratio value sorted from large to small as operators according to a fitness ratio method for the rest individuals, namely selecting the probability that the fitness of the individual accounts for the fitness of all the rest individuals:
Figure FDA0003082906020000061
wherein: p is the probability of selecting the individual as an operator; q is fitness;
carrying out two-group random matching on the selected 20 operators to obtain 10 operator pairs; performing cross operation on each group of operator pairs in a consistent hybridization mode until 150 new individuals are generated; performing mutation operation on 150 new individuals, firstly setting the mutation probability to be 0.05, selecting the individuals to be mutated, randomly selecting mutation sites for performing real value mutation on the individuals to be mutated, and replacing the new individuals subjected to the mutation operation with the individuals to be mutated to obtain 150 offspring individuals; adding the 150 sub-generation individuals and the individuals with the fitness of the first 25 percent in the first generation population to generate a second generation population; the third generation population is obtained on the basis of the second generation population according to the generation process of the second generation population; the objects of the cross operation and the mutation operation are variable parameters in a group of input, the fixed parameters and the characterization parameters are fixed values, and the fixed parameters and the characterization parameters only relate to the calculation of individual fitness;
(4) after each new generation of population is generated, determining whether to continuously update the population by iteration or terminate iteration according to an iteration termination condition; the termination conditions are specifically:
1) the preset termination fitness is set as QmSetting the maximum fitness of the nth generation population as QnN is not less than 2, and Q is judgednWhether or not it is greater than QmIf yes, executing step 2), otherwise, executing step 3);
2) presetting a fitness amplification threshold value as eta, and judging QnComparison of Qn-1If the amplification is less than eta, if so, the iteration is terminated, and Q and the output are outputnCorresponding individuals, otherwise, executing the step 3);
3) the threshold value of the maximum algebra of the preset population is GmJudging whether the current population algebra reaches the maximum algebra, if so, terminating iteration and outputting the maximum fitness Q in the current populationnOtherwise, generating an n +1 generation population according to the step (4) on the basis of the n generation population, skipping to the step 1), and enabling the maximum fitness Q of the n +1 generation populationn+1And QmPerforming size judgment to determine whether to execute the step 2) or the step 3) next;
(5) decoding the individual corresponding to the maximum fitness in the population meeting the iteration termination condition in the step (4) to obtain the optimal value of each index of the variable parameter, namely the optimal value of each index of the material attribute;
(6) setting the material attribute as a fixed parameter, wherein the value of the fixed parameter combination is the optimal value of each index of the material attribute obtained in the step (5); and (5) setting the mechanical index as a variable parameter, keeping the representation parameter unchanged, and repeating the steps (2) to (5) to obtain the optimal value of the mechanical index.
2. The artificial intelligence-based road pavement full life cycle quality tracing method according to claim 1, wherein in the first step, the mechanical indexes of the construction period data comprise: 1. the roadbed deflection, namely the rebound deflection at different positions is measured by adopting a Beckman beam method; 2. the compactness of each layer specifically comprises roadbed compactness, water stability lamination compactness and pavement compactness; 3. the cubic compressive strength of cement mortar specifically comprises the average values of the breaking load and the compressive strength of the cubic body made of the cement mortar at different mixing ratios of each key part.
3. The artificial intelligence-based road pavement full life cycle quality tracing method according to claim 1, wherein in the first step, the material attribute indexes of the construction period data comprise: 1. the soil sample screening results comprise coarse screening results and fine screening results, wherein the coarse screening results are the percentage of the total soil mass under each large aperture, and the large apertures are 2mm, 5mm and 10mm respectively; the result of the fine screening is the percentage of the total soil mass under each small aperture, and the small apertures are 1mm, 0.5mm, 0.25mm and 0.075mm respectively; 2. the glue-sand ratio of the water stable layer; 3. the California bearing ratio of the roadbed is the ratio of the unit pressure when the penetration of the roadbed soil reaches 2.5mm to the standard load when the standard crushed stone is pressed into the roadbed with the same penetration.
4. The artificial intelligence-based road pavement full life cycle quality tracing method according to claim 1, wherein in the first step, the traffic condition indexes of the operation period data comprise: 1. the traffic axle load is that the traffic volume is converted into equivalent traffic axle load according to the type of traffic unit; 2. the passenger-cargo ratio can assist the equivalent axle load to distinguish the influence of different types of vehicles on the road surface damage.
5. The artificial intelligence-based road pavement full life cycle quality tracing method according to claim 1, wherein in the first step, the environmental factor indexes of the operation period data comprise: 1. hydrological factors mainly including the influence of precipitation and humidity environment; 2. atmospheric factors including atmospheric pressure, wind speed and temperature in the particular area.
6. The artificial intelligence-based full life cycle quality tracing method for road surfaces as claimed in claim 1, wherein in the first step, the indexes of the service performance data comprise road surface damage condition, road surface running quality, road surface rutting, road surface skid resistance, road surface structural strength, road surface service performance and road surface damage reduced area.
7. The artificial intelligence-based highway pavement full life cycle quality tracing method according to claim 1, wherein the concrete maintenance decision making process in the fourth step is as follows: summarizing the predicted data of the service performance data in the accumulated period characteristic parameters of all hectometer pile numbers of the target time node, giving out performance weight and grading basis according to the local expressway condition, weighting all service performance to obtain the comprehensive performance indexes of all hectometer pile numbers, grading the comprehensive performance of different pile numbers according to the grading basis, selecting the maintenance standard grade, wherein the pile numbers under the grade are the maintenance pile numbers, and making a decision according to each refined service performance by using specific maintenance measures.
CN202110571984.9A 2021-05-25 2021-05-25 Highway pavement full life cycle quality tracing method based on artificial intelligence Active CN113177735B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110571984.9A CN113177735B (en) 2021-05-25 2021-05-25 Highway pavement full life cycle quality tracing method based on artificial intelligence

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110571984.9A CN113177735B (en) 2021-05-25 2021-05-25 Highway pavement full life cycle quality tracing method based on artificial intelligence

Publications (2)

Publication Number Publication Date
CN113177735A CN113177735A (en) 2021-07-27
CN113177735B true CN113177735B (en) 2022-03-08

Family

ID=76928110

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110571984.9A Active CN113177735B (en) 2021-05-25 2021-05-25 Highway pavement full life cycle quality tracing method based on artificial intelligence

Country Status (1)

Country Link
CN (1) CN113177735B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113918538B (en) * 2021-10-14 2024-05-14 河北工业大学 New road maintenance data migration system based on artificial neural network
CN117273477A (en) * 2023-08-16 2023-12-22 湖北交投智能检测股份有限公司 Asphalt highway pavement maintenance decision method, device, equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993223A (en) * 2019-03-26 2019-07-09 南京道润交通科技有限公司 Pavement Condition prediction technique, storage medium, electronic equipment
CN111105332A (en) * 2019-12-19 2020-05-05 河北工业大学 Highway intelligent pre-maintenance method and system based on artificial neural network
CN112733442A (en) * 2020-12-31 2021-04-30 交通运输部公路科学研究所 Road surface long-term performance prediction model based on deep learning and construction method thereof

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993223A (en) * 2019-03-26 2019-07-09 南京道润交通科技有限公司 Pavement Condition prediction technique, storage medium, electronic equipment
CN111105332A (en) * 2019-12-19 2020-05-05 河北工业大学 Highway intelligent pre-maintenance method and system based on artificial neural network
CN112733442A (en) * 2020-12-31 2021-04-30 交通运输部公路科学研究所 Road surface long-term performance prediction model based on deep learning and construction method thereof

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
高等级沥青路面使用性能预测模型及预防护性养护措施研究;侯超平;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20160415;C034-44 *

Also Published As

Publication number Publication date
CN113177735A (en) 2021-07-27

Similar Documents

Publication Publication Date Title
Li et al. Automated decision making in highway pavement preventive maintenance based on deep learning
Piryonesi et al. Role of data analytics in infrastructure asset management: Overcoming data size and quality problems
CN113177735B (en) Highway pavement full life cycle quality tracing method based on artificial intelligence
Hossain et al. Artificial neural network modelling to predict international roughness index of rigid pavements
Kayadelen Soil liquefaction modeling by genetic expression programming and neuro-fuzzy
Kayadelen et al. Modeling of the angle of shearing resistance of soils using soft computing systems
Kiran et al. Shear strength prediction of soil based on probabilistic neural network
Domitrović et al. Application of an artificial neural network in pavement management system
Qasem et al. Optimizing ANFIS for sediment transport in open channels using different evolutionary algorithms
CN109102016B (en) A kind of test method for bridge technology situation
Liu et al. Optimizing asphalt mix design through predicting the rut depth of asphalt pavement using machine learning
Ghajar et al. An adaptive network-based fuzzy inference system for rock share estimation in forest road construction
CN114912635A (en) Highway maintenance decision-making method and system, storage medium and electronic equipment
Al-Busultan et al. Application of artificial neural networks in predicting subbase CBR values using soil indices data
Wang et al. Building an improved artificial neural network model based on deeply optimizing the input variables to enhance rutting prediction
Baldo et al. Road pavement asphalt concretes for thin wearing layers: a machine learning approach towards stiffness modulus and volumetric properties prediction
Hao et al. Prediction of CBR and resilient modulus of crushed waste rocks using machine learning models
CN115948964A (en) Road flatness prediction method based on GA-BP neural network
Othman et al. Prediction of the optimum asphalt content using artificial neural networks
Kim et al. Optimizing the mixing proportion with neural networks based on genetic algorithms for recycled aggregate concrete
Mahmood et al. A unified artificial neural network model for asphalt pavement condition prediction
CN117350146A (en) GA-BP neural network-based drainage pipe network health evaluation method
Yang et al. Forecasting model for urban traffic flow with BP neural network based on genetic algorithm
Saltan et al. Comparative analysis of using artificial neural networks (ANN) and gene expression programming (GEP) in backcalculation of pavement layer thickness
Qian Road pavement performance evaluation model based on hybrid genetic algorithm neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant