An intelligent computing and scheme generating system for road construction material proportioning

Abstract

The application relates to the technical field of computer aided engineering and material science, and discloses a pavement construction material proportioning intelligent calculation and scheme generation system. The system firstly constructs a response manifold model under multi-scale constraints, accurately characterizes a nonlinear mapping relationship between material proportioning parameters, working condition disturbance parameters and macroscopic performance indexes; utilizes an active learning and model iteration module to iteratively optimize the model in a manner of minimizing the number of physical tests; subsequently, a proportioning scheme reverse generation module reversely solves a robust benchmark proportioning scheme which satisfies performance constraints and is insensitive to working condition disturbances based on the model; and finally, a construction process self-adaptive correction module acquires working condition disturbance data in real time in a construction stage, calculates proportioning adjustment amounts by using the model, and realizes dynamic correction of a production process. The application significantly improves the efficiency and quality of proportioning design, ensures the robustness of a proportioning scheme, and realizes closed-loop control of a design stage and a production stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided engineering and material science, in particular to an intelligent calculation and scheme generation system for pavement construction material proportioning. BACKGROUND

[0002] The mix design of pavement construction materials, especially asphalt mixtures and cement stabilized macadam, is a key link to determine the final performance, durability and engineering cost of the pavement.

[0003] Currently, the mix design process of pavement materials largely relies on empirical trial mixing and orthogonal test methods in the laboratory. This traditional design paradigm first requires technicians to preset a wide range of parameters based on engineering specifications and personal experience, and then explores the material performance under different mix combinations through a large number of physical tests. This process is essentially a trial-and-error method, which not only consumes a large amount of raw materials and human resources, but also has a long test cycle, directly affecting the efficiency of new material research and development and the progress of engineering.

[0004] In addition, the so-called optimal mix found under the idealized and constant conditions in the laboratory often shows instability in actual engineering applications. This is because the traditional design method usually fails to fully consider the coupling effects of inherent fluctuations between batches of raw materials, dynamic changes in the construction site environment (such as temperature and humidity), and other uncertain disturbance factors on the final performance. Therefore, a mix scheme that performs well on laboratory data often lacks robustness and its performance may deteriorate significantly when faced with complex real-world working conditions, making it difficult to ensure the consistency of the final product quality.

[0005] Furthermore, the existing technology is significantly disconnected between the design stage and the production stage. The static mix scheme determined in the laboratory is directly applied to the dynamically changing production site, lacking an effective closed-loop feedback and real-time correction mechanism. When disturbances occur during the construction process, such as an increase in aggregate moisture content due to weather changes, the on-site response measures rely on the experience and judgment of construction personnel for temporary adjustments. This adjustment lacks precise quantitative basis and it is difficult to ensure that the adjusted mix can still maintain the performance indicators within the optimal range, which poses a risk to the final pavement quality. At the same time, the design data and test experience accumulated from a large number of past engineering projects have not been systematically utilized to provide effective knowledge support for the design of new projects. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an intelligent calculation and scheme generation system for pavement construction material proportioning, which solves the problems of low efficiency, poor robustness of traditional pavement material mix design, and the disconnection between the design scheme and the dynamically changing actual construction process.

[0007] To achieve the above object, the present application is implemented by the following technical solutions: the present application provides a road construction material proportioning intelligent calculation and scheme generation system.

[0008] The system introduces a response manifold model capable of fusing historical engineering experience and multi-scale physical constraints, and establishes an inverse solving path from performance requirements to robust proportioning scheme and a dynamic closed-loop correction mechanism from the design end to the construction end based on the model, thereby systematically solving the above defects of the prior art.

[0009] In one specific embodiment, the system comprises:

[0010] A priori knowledge transfer module, which functions to retrieve and select one or more historical models from a historical manifold model library according to the input new project meta-information, and outputs the parameters of the selected model as a priori model parameters;

[0011] A response manifold construction module under multi-scale constraints, which functions to receive the a priori model parameters and construct a multi-dimensional response manifold model representing the mapping relationship between material proportioning, working condition disturbance and macroscopic performance;

[0012] A proportioning scheme inverse generation module, which functions to receive the multi-dimensional response manifold model and an external user input performance requirement package, the performance requirement package containing constraints on material performance, and performs inverse solving calculation to determine a benchmark proportioning scheme satisfying the performance constraints in the performance requirement package;

[0013] An active learning and model iteration module, which functions to analyze the uncertainty of the multi-dimensional response manifold model and calculate the parameter point with the maximum model prediction variance, which is recommended for subsequent physical tests, and the results of the physical tests are fed back to establish new data points and iteratively update the multi-dimensional response manifold model;

[0014] A construction process adaptive correction module, which functions to receive real-time monitored construction site disturbance parameters according to the benchmark proportioning scheme output by the proportioning scheme inverse generation module, and dynamically adjusts the material proportioning according to the actual measurement data and the multi-dimensional response manifold model.

[0015] Further, the a priori knowledge transfer module is internally configured with a historical manifold model library unit and a similarity evaluation and model selection unit. The historical manifold model library unit is used to structurally store the response manifold model parameters of historical engineering projects and the corresponding project meta-information. When a new project is started, the similarity evaluation and model selection unit calculates the similarity between the input new project meta-information vector and each historical project meta-information vector stored in the library. The calculation can use the following formula:

[0016] ;

[0017] wherein, is a similarity score, is the dimension of the meta-information vector, is the preset weight of the th feature, and are the th feature value of the new item and the historical item respectively, is a distance function. The unit selects the historical model with the highest similarity score and extracts its model parameters as the prior knowledge for the training of the new model, is a distance function for calculating the distance between two feature values. When the th feature is a categorical feature, if and are the same, the distance function takes a value of 0, otherwise takes a value of 1. When the th feature is a numerical feature, the distance function calculates the absolute difference between the normalized values.

[0018] Further, the response manifold construction module under the multi-scale constraint is internally configured with a parameter space definition unit, a microstructure constraint unit and a manifold modeling and training unit. The parameter space definition unit is used to configure the value range of the material ratio parameter vector and the working condition disturbance parameter vector . The microstructure constraint unit integrates one or more physical model analysis tools, such as a discrete element method simulator, for analyzing the micro behavior of materials under a given ratio and outputting a set of micro physical indicators . The manifold modeling and training unit constructs a multi-dimensional response manifold model using Gaussian process regression, assuming that the performance vector obeys a Gaussian process distribution with the parameter vector as input:

[0019] ;

[0020] wherein, is a mean function, whose initial form is set by the prior model parameters output by the prior knowledge transfer module, is a covariance kernel function, denotes a Gaussian process. The training process of this unit aims to minimize a composite objective function :

[0021] ;

[0022] wherein, Based on experimental datasets The negative logarithmic marginal likelihood term, This is a physical constraint penalty term. It applies to the physical properties output by the microstructure constraint unit. When the value exceeds the preset reasonable range, this item incurs a penalty. where represents the weighting coefficients. This composite objective function ensures that the final manifold model can both fit the macroscopic experimental data and follow fundamental physical laws.

[0023] Furthermore, the reverse generation module for the proportioning scheme is internally configured with a requirement definition unit, a feasible region location unit, and a robust optimization unit. The requirement definition unit is used to receive and parse the performance requirement package, transforming it into a mathematical set of performance constraints. The feasible region location unit searches for all feasible parameter sets that satisfy performance constraints within the matching parameter space by slicing the constructed response manifold model. Robust optimization unit in feasible parameter set Internally, seeking a way to deal with disturbances The least sensitive optimal ratio scheme The goal of this optimization process is to minimize the average gradient norm of performance with respect to perturbations:

[0024] ;

[0025] in, This is the optimal robust ratio scheme for the final output. It is the set of feasible parameters determined by the feasible region location unit. The operating condition disturbance parameter vector Prior assumptions about the probability distribution, such as a baseline perturbation. A Gaussian distribution centered on the center, Indicates the impact of disturbances Find the expected value of all possible values. It is the mean function of the response manifold model. For the perturbation vector The gradient, and the L2 norm of the gradient vector. The square of the quantification is the ratio of Disturbance is Under certain conditions, the instantaneous sensitivity of macroscopic performance to disturbances is determined. By solving this optimization problem, the unit outputs a unique baseline configuration. .

[0026] Furthermore, the active learning and model iteration module is internally configured with an uncertainty quantification unit and a test point recommendation unit. The uncertainty quantification unit is used to extract the prediction variance from the Gaussian process regression model. , which quantifies the uncertainty of the model prediction at the parameter point . The test point recommendation unit determines the next most information gain test point by solving

[0027] ;

[0028] wherein, represents the parameter point maximizing the prediction variance, the next most information gain test point is determined, thereby efficiently guiding the test and iteratively updating the model using the newly added test data.

[0029] Further, the construction process adaptive correction module is internally configured with a data acquisition unit, a real-time performance evaluation unit and a correction decision unit. The data acquisition unit is used to monitor and collect key disturbance parameters such as temperature, humidity, etc. in the construction site in real time, and obtain the actual disturbance vector . The real-time performance evaluation unit inputs the current proportioning and the actual disturbance according to the trained multi-dimensional response manifold model, and calculates the performance prediction value of the material under the current working condition in real time. The correction decision unit calculates the material proportioning amount that needs to be adjusted based on the error between the current material performance prediction value and the performance constraints in the performance requirement package. This calculation is completed by solving a linear equation system:

[0030] ;

[0031] wherein, and are the Jacobian matrices of the response manifold model with respect to the proportioning and disturbance at the current working point, is the deviation between the actual disturbance and the reference disturbance. Before outputting the optimized material proportioning, the correction decision unit can also perform a verification step, inputting the optimized proportioning, calling the real-time performance evaluation unit to predict its expected performance, and judging whether the expected performance meets the constraints in the performance requirement package to ensure the effectiveness of the correction.

[0032] The present application provides a pavement construction material proportioning intelligent calculation and scheme generation system. It has the following beneficial effects:

[0033] 1. The present application sets up an active learning and model iteration module, which can intelligently recommend the test point with the highest information value based on the uncertainty quantification of the model, thereby avoiding a large number of blind physical tests, greatly reducing the test workload and material consumption, significantly improving the efficiency of proportioning design and reducing the research and development cost, and accelerating the research and development cycle of new materials.

[0034] 2. The multidimensional response manifold model constructed in this invention unifies the modeling of material properties, proportioning parameters, and operating condition disturbance parameters. Through the robust optimization unit in the proportioning scheme reverse generation module, it can find the optimal solution least sensitive to uncertainties such as raw material fluctuations and changes in the construction environment when solving for proportioning schemes that meet performance requirements. This makes the final generated proportioning scheme more stable and consistent in practical applications, improving the robustness and reliability of the proportioning scheme.

[0035] 3. The adaptive correction module for the construction process of this invention can monitor on-site disturbance parameters in real time and quickly calculate the optimal mix adjustment amount using the gradient information of the trained response manifold model. This achieves closed-loop control from static design in the laboratory to dynamic production, seamlessly connecting precision design with the complex on-site production process, forming an intelligent production system that can respond in real time and self-correct, effectively ensuring the quality of the final product. Attached Figure Description

[0036] Figure 1 This is a block diagram of the system functional modules of the present invention;

[0037] Figure 2 This is a diagram showing the internal structure and data flow of the prior knowledge transfer module of this invention.

[0038] Figure 3 This is a diagram showing the internal structure and data flow of the response manifold construction module under multi-scale constraints of the present invention.

[0039] Figure 4 This is a diagram showing the internal structure and data flow of the reverse generation module for the proportioning scheme of the present invention;

[0040] Figure 5 This is a diagram showing the internal structure and data flow of the active learning and model iteration module of this invention;

[0041] Figure 6 This is a diagram showing the internal structure and data flow of the adaptive correction module for the construction process of the present invention. Detailed Implementation

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] See attached document Figure 1 This invention provides an intelligent calculation and scheme generation system for road construction material proportions, which can run on one or more computing devices.

[0044] In one specific embodiment, the system's operating environment can include: a central processing server for performing main computing tasks, model training and optimization solving; a data storage device, such as a database server, for storing historical manifold model library, project test data, trained model parameters and generated proportioning scheme; one or more user interaction terminals for technicians to input project meta-information, performance requirement package and view recommended test points and proportioning schemes output by the system; and data acquisition devices, such as temperature sensors, humidity sensors and material property monitors, deployed at the construction site for collecting real-time working condition disturbance parameters.

[0045] The central processing server, data storage device and user interaction terminal can communicate data through a local area network or a wide area network. The data acquisition device can send the collected data to the central processing server in real time through a wired or wireless communication network.

[0046] As shown in Figure 1 The system provided by the present application includes the following software functional layers: prior knowledge transfer module, response manifold construction module under multi-scale constraints, proportioning scheme reverse generation module, active learning and model iteration module, and construction process adaptive correction module.

[0047] The function of the prior knowledge transfer module is to respond to a new project start request. It receives the new project meta-information input by the user interaction terminal, accesses the historical manifold model library stored in the data storage device, and outputs a prior model parameter after evaluation and calculation.

[0048] The response manifold construction module under multi-scale constraints receives the prior model parameter and combines the initial test data set obtained from the data storage device. The core function of this module is to construct and train a multi-dimensional response manifold model. This model represents the nonlinear mapping relationship between material proportioning parameters, working condition disturbance parameters and macroscopic performance indicators. The trained multi-dimensional response manifold model is stored and provided for subsequent module calls.

[0049] The proportioning scheme reverse generation module receives the multi-dimensional response manifold model constructed by the response manifold construction module under multi-scale constraints, and receives a performance requirement package from the user interaction terminal. This module performs reverse solving calculation based on the model to determine a benchmark proportioning scheme that meets all the constraints in the performance requirement package, and outputs the scheme to the user interaction terminal or the data storage device.

[0050] The active learning and model iteration module also receives the multi-dimensional response manifold model. This module analyzes the prediction uncertainty, such as the prediction variance, of the model and calculates a recommended test point parameter. This parameter point is sent to the user interaction terminal to guide the technician to conduct the next physical test. When the new test data point, including the test parameter and the test result, is input into the system, the new data point is added to the test data set and triggers the response manifold construction module under multi-scale constraints to iteratively update the multi-dimensional response manifold model.

[0051] The construction process adaptive correction module is activated during the construction phase. It first obtains the determined reference mix design from the mix design inverse generation module. At the same time, it receives the actual measurement values of the site disturbance parameters in real time through the data acquisition device. Based on the multi-dimensional response manifold model, the reference mix design, and the actual disturbance values, this module calculates a mix adjustment amount and generates a corrected mix instruction. This instruction can be sent to the control system of the mixing plant at the construction site to achieve dynamic adjustment of the production mix.

[0052] Referring to the drawings Figure 2 The function of the prior knowledge transfer module is to determine an initial predictive prior model for the newly started project by using historical engineering data, thereby providing a high starting point for subsequent response manifold construction.

[0053] In a specific embodiment, the module includes a historical manifold model library unit and a similarity evaluation and model selection unit.

[0054] The historical manifold model library unit is a structured database deployed on the data storage device. This database is used to persistently store the data and models of multiple completed engineering projects in the past. For each historical project in the library, the unit stores an independent record. Each record contains the following fields: project meta information, which is a vector describing the objective characteristics of the project; test data set, which is the data point set of all physical tests performed during the design process of the project; and manifold model parameters, which are all parameters of the Gaussian process regression model finally trained based on the test data of the project, including the type of the selected kernel function, the optimized kernel function hyperparameters, and the noise variance.

[0055] The similarity evaluation and model selection unit is a calculation unit that executes a determined process to select the optimal prior model. This unit first receives the meta information vector describing the characteristics of the current new project input from the user interaction terminal After receiving , the unit accesses the historical manifold model library unit and traverses each historical project record stored in the library to extract the meta information vector of each historical project. For each , the unit calculates its similarity score with . The calculation is performed by the following formula:

[0056] ;

[0057] where is the similarity score, is the dimension of the meta-information vector, is the preset weight of the th feature, and are the th feature values of the new item and the historical item, respectively, is the distance function. The unit selects the historical model with the highest similarity score and extracts its model parameters as the prior knowledge for the training of the new model, is a distance function for calculating the distance between two feature values. When the th feature is a categorical feature, if and are the same, the distance function takes the value of 0, otherwise it takes the value of 1. When the th feature is a numerical feature, the distance function calculates the absolute difference between the normalized values.

[0058] After the similarity score calculation of all historical items in the library is completed, the unit determines the index of the historical item with the highest score.

[0059] Finally, the unit retrieves the corresponding manifold model parameters from the historical manifold model library unit according to the index and extracts the trained mean function from it. The mean function is output as the prior model parameter and passed to the response manifold construction module under multi-scale constraints for use as the prior mean function for the training of the new model.

[0060] Referring to the accompanying drawings Figure 3 , the function of this module is to receive prior model parameters and initial experimental data and construct a multi-dimensional response manifold model that can accurately represent the nonlinear mapping relationship between material proportioning, working condition disturbance, and macroscopic performance.

[0061] In a specific embodiment, the module includes a parameter space definition unit, a microstructure constraint unit, and a manifold modeling and training unit

[0062] The parameter space definition unit is used to structurally define all variables involved in modeling. The unit unifies all input variables into a parameter vector ​, which consists of two parts: a controllable material proportioning parameter vector , whose components represent the precisely controllable process parameters such as aggregate proportion, asphalt-aggregate ratio, and admixture dosage, and an uncontrollable working condition disturbance parameter vector , whose components represent the uncontrollable disturbance factors such as batch fluctuation of raw material physical indexes and construction environment temperature and humidity. This unit also defines the output of the model, i.e., a macroscopic performance vector , whose components represent the performance indexes that need to be predicted, such as high-temperature dynamic stability, low-temperature cracking resistance, and water stability.

[0063] The microstructure constraint unit integrates a physical model analysis tool, such as a discrete element method simulator. The function of this unit is to impose physical constraints on the construction of the response manifold model to ensure the physical authenticity of the model prediction. This unit receives the material proportioning vector output by the parameter space definition unit as input and calls the internal physical model analysis tool to perform a mapping calculation , thereby obtaining a physical index vector characterizing the state of the material microstructure. The physical indexes can include the locking index of the aggregate skeleton, the asphalt-aggregate interface energy, and the uniformity of the void ratio distribution. The unit internally presets reasonable value intervals for these physical indexes, constituting a physically feasible domain .

[0064] The manifold modeling and training unit is the core calculation unit for model construction and optimization. This unit uses Gaussian process regression (GPR) as the modeling technique and assumes that the performance vector obeys a Gaussian process distribution with the parameter vector as input:

[0065] ;

[0066] where is the mean function, whose initial form is set by the prior model parameters output by the prior knowledge transfer module, is the covariance kernel function, and represents the Gaussian process. The training process of this unit aims to minimize a composite objective function by optimizing a set of hyperparameters of the kernel function and the model noise variance. The composite objective function is composed of a data fitting term and a physical constraint term:

[0067] ;

[0068] in, Based on experimental datasets The negative logarithmic marginal likelihood term, It is a physical constraint penalty term, which is applied when the physical properties output by the microstructure constraint unit are... When the value exceeds the preset reasonable range, this item incurs a penalty. The weighting coefficient is determined by the proportion. Calculated microphysical index vector The component exceeds the preset physical feasible domain. hour, Take a positive value, otherwise zero. λ is a weighting coefficient used to balance data fitting and physical constraints. By minimizing this composite objective function, the unit ultimately generates an optimized multidimensional response manifold model that simultaneously satisfies macroscopic data fitting and microscopic physical laws.

[0069] See attached document Figure 4 The function of this module is to receive the constructed multidimensional response manifold model and the performance requirements of external input, and perform inverse solution calculations to determine a baseline mix design that simultaneously meets performance constraints and is insensitive to operating condition disturbances.

[0070] In one specific implementation, the module includes a requirement definition unit, a feasible domain location unit, and a robust optimization unit.

[0071] The requirements definition unit receives performance requirement packages input from the user interface terminal. These packages contain one or more specific requirements for the macroscopic properties of the material. The function of this unit is to parse and transform these performance requirements, described in engineering language, into a set of deterministic mathematical constraints. For example, a requirement for a high-temperature dynamic stability of not less than 5000 cycles / mm is transformed into… ,in It is a performance vector The component corresponding to dynamic stability. This unit combines all the analyzed requirements into a performance constraint set. And output it.

[0072] The feasible region localization unit receives the multidimensional response manifold model output by the response manifold construction module under multi-scale constraints, and the performance constraint set output by the requirement definition unit. This unit searches for all proportioning parameters that satisfy the performance constraints within the material proportioning parameter space. Specifically, for each candidate proportioning vector in the proportioning parameter space, this unit... Under the set reference disturbance conditions Next, call the mean function of the response manifold model. Predict its performance vector .like all components of the performance constraint set if the corresponding constraint in is satisfied, the recipe vector is determined as feasible. The unit collects all feasible recipe vectors determined as feasible and forms a feasible recipe set .

[0073] The robustness optimization unit receives the feasible recipe set and the multi-dimensional response manifold model. The function of the unit is to further search for an optimal solution that is least sensitive to the operating condition perturbation from all recipe solutions that satisfy the basic performance requirement. The search is realized by solving an optimization problem whose objective is to minimize the expected value of the gradient norm of the performance with respect to the perturbation parameters. The optimization problem is formulated as follows:

[0074] ;

[0075] where, is the optimal robust recipe solution as the final output, is the feasible recipe set determined by the feasible region localization unit, is the prior assumption of the probability distribution of the operating condition perturbation parameter vector , for example, a Gaussian distribution centered at the baseline perturbation , denotes the mathematical expectation over all possible values of the perturbation , is the gradient of the mean function of the response manifold model with respect to the perturbation vector , the L2 norm of the gradient vector quantifies the instantaneous sensitivity of the macroscopic performance to the perturbation under the condition that the recipe is and the perturbation is . By solving this optimization problem, the unit outputs a unique baseline recipe solution .

[0076] Referring to the attached drawings, Figure 5 the function of the module is to analyze the prediction uncertainty of the current multi-dimensional response manifold model in the entire parameter space and recommend a test point with the most information gain based on this to guide the subsequent physical test, and finally use the new test data to iteratively optimize the model.

[0077] In a specific embodiment, the module includes an uncertainty quantification unit, a test point recommendation unit, and a data feedback and model updating unit.

[0078] The uncertainty quantification unit receives the current trained multi-dimensional response manifold model outputted by the response manifold construction module under multi-scale constraints. Since the model is built based on Gaussian Process Regression (GPR), the unit can directly extract the prediction variance at any unobserved parameter point from the mathematical properties of the GPR model . The prediction variance is used as a quantitative indicator of the uncertainty of the model's prediction at the point. The greater the value of the prediction variance, the less sufficient the model's cognition is in the point and its vicinity.

[0079] The test point recommendation unit receives the function of the prediction variance outputted by the uncertainty quantification unit. The unit performs a global optimization search within the variable value range defined by the parameter space definition unit to find the parameter point that maximizes the prediction variance. The search process is achieved by solving the following optimization problem: The obtained is the next recommended physical test point. The parameter point is then sent to the user interaction terminal for the technical staff to perform the corresponding physical test.

[0080] The data feedback and model updating unit is used to receive a newly generated test data point. When the technical staff completes the physical test according to the recommended test point , the test point and its corresponding test result are inputted into the system through the user interaction terminal as a new data record. The unit adds the new data record to the test data set in the data storage device.

[0081] After adding the new data, the unit immediately triggers the response manifold construction module under multi-scale constraints to retrain and iteratively update the multi-dimensional response manifold model based on the expanded test data set. This process constitutes a closed loop of model prediction, physical test, and model updating, so that the model is prioritized to be improved in the area with the greatest uncertainty.

[0082] Referring to the accompanying Figure 6 , the function of this module is to dynamically adjust the reference proportioning scheme in real time in response to the disturbance of the field working conditions during the actual production process of pavement materials, to ensure that the performance of the final produced materials always meets the preset performance requirements.

[0083] In a specific embodiment, the module includes a data acquisition unit, a real-time performance evaluation unit, and a correction decision unit.

[0084] ​The data acquisition unit connects to multiple sensors and monitoring devices deployed at the construction site via a communication interface. These devices are used to monitor key operational disturbance parameters in real time, and the monitored data includes, but is not limited to, ambient temperature, ambient humidity, and physical properties of raw materials (such as aggregates). The unit collects this data at a preset time frequency (e.g., once per minute) and integrates it into a real-time operational disturbance vector. This vector is then passed to the real-time performance evaluation unit and the correction decision unit.

[0085] The real-time performance evaluation unit receives the baseline proportioning scheme determined by the proportioning scheme reverse generation module. and the real-time operating condition disturbance vector provided by the data acquisition unit. The core function of this unit is to invoke a trained multidimensional response manifold model. It combines vectors... As input, the mean function of the response manifold model is substituted. Thus, the predicted values ​​of the material's macroscopic properties under the combined effect of the baseline mix ratio and actual disturbances can be calculated. .

[0086] The correction decision unit is the core computational unit that performs ratio adjustments. This unit first receives the performance prediction values ​​output by the real-time performance evaluation unit. And obtain the set of performance constraints in the performance requirement package from the reverse generation module of the ratio scheme. It calculates With performance constraints The error between them.

[0087] To eliminate this error, this unit calculates the proportion adjustment based on the response manifold model. This calculation process includes: obtaining the baseline ratio. The corresponding reference perturbation vector And calculate the deviation between the current actual disturbance and the reference disturbance. Subsequently, the unit operates at the current work point. At this point, calculate the Jacobian matrix of the mean function of the response manifold model with respect to the matching parameters. and the Jacobian matrix with respect to the perturbation parameters Finally, this unit determines the optimal proportion adjustment by solving the following system of linear equations. :

[0088] ;

[0089] in, and These are the Jacobian matrices of the response manifold model at the current operating point for the ratio and perturbation. is the deviation between the actual disturbance and the reference disturbance, the solution of the system of equations i.e. the current disturbance is offset. The unit adds this adjustment to the reference mix proportion .

[0090] Before outputting the optimized mix proportion, the revision decision unit further performs a validation step. The unit takes the optimized mix proportion and the current real-time operating disturbance vector as inputs, and calls the real-time performance evaluation unit again. The real-time performance evaluation unit returns the expected performance under the optimized mix proportion. The revision decision unit determines whether all the performance constraints in the performance requirement envelope are satisfied. If yes, the unit outputs the optimized mix proportion (or the adjustment ) as the revision instruction to the central control system of the mixing plant at the construction site, to perform real-time update of the mix proportion.

[0091] The complete workflow of the system is illustrated in detail below, taking the intelligent design and production of a high-performance asphalt mixture (e.g. AC-20 type) as an example.

[0092] First, a new project is initiated. The technical staff inputs the key meta-information of the project through the user interaction terminal, e.g. the material type is AC-20 heavy traffic asphalt mixture, and the application operating condition is high-temperature and rainy area. The prior knowledge transfer module receives the new project meta-information vector and accesses the historical manifold model library unit. The module retrieves one or more historical items from the library by performing similarity calculation, e.g. an AC-25 high-temperature area project, and determines that it has the highest similarity score. Subsequently, the module extracts the manifold model parameters of the historical project and outputs them as the prior model parameters.

[0093] Next, the system enters the model construction and iteration phase. The response manifold construction module under multi-scale constraints receives the above-mentioned prior model parameters. The technical staff sets the variable range of this design through the parameter space definition unit: the material mix proportion vector including the oil-aggregate ratio, the proportion of each aggregate grade, and the filler content; the operating disturbance vector including the mixing temperature and the aggregate moisture content; the macroscopic performance vector including the high-temperature dynamic stability, the low-temperature crack resistance, and the water stability residual splitting tensile strength (TSR). At the same time, the microstructure constraint unit is configured as a physical model for calculating the void ratio (VMA) of asphalt mixture, and sets a reasonable interval of VMA (e.g. 13% to 15%) as the physical constraint.

[0094] The system utilizes a very small initial experimental data set (e.g. 5 groups of historical data or orthogonal experimental points) and the above-mentioned physical constraints to train an initial multi-dimensional response manifold model through manifold modeling and training unit.

[0095] Subsequently, the active learning and model iteration module intervenes. The uncertainty quantification unit analyzes the prediction variance of the initial model and finds that the model has the highest uncertainty in a certain untested parameter region (e.g. the combination of high oil-stone ratio and low filler content). The experimental point recommendation unit determines this parameter point as the first recommended experimental point and sends it to the user terminal.

[0096] The laboratory technician performs physical tests (e.g. performs rutting tests and splitting tests) according to the parameter configuration and inputs the actual performance results measured by the tests into the terminal. The data feedback and model updating unit adds this new data point to the experimental data set. This triggers the response manifold construction module under multi-scale constraints to retrain and update the multi-dimensional response manifold model based on the expanded data set. This recommended-test-feedback-update closed loop process is repeated several times (e.g. 10 to 15 times) until the average prediction variance of the entire model is below the preset convergence threshold. At this point, a high-precision multi-dimensional response manifold model is constructed.

[0097] In the last step of the design phase, the technician inputs the final performance requirement package, e.g. {dynamic stability} 5000 times / mm, TSR ≥ 85%. The mixture proportion scheme reverse generation module receives this requirement package and the trained mature response manifold model. The feasible region positioning unit first searches for all mixture proportion schemes that satisfy the above performance constraints under the set reference working condition disturbance (e.g. aggregate moisture content 4.0%). Then, the robustness optimization unit solves the robustness optimization objective based on the expected fluctuation distribution of working condition disturbance (particularly aggregate moisture content) in the feasible set to finally determine an optimal mixture proportion scheme that is least sensitive to moisture content fluctuations . This is determined as the reference mixture proportion scheme.

[0098] Finally, in the construction production phase, the reference mixture proportion scheme is loaded into the construction process self-adaptive correction module deployed at the mixing plant. The data acquisition unit connects with the moisture sensor on the aggregate conveyor belt to monitor the current aggregate moisture content in real time, which is 4.6%. The correction decision unit immediately calculates the disturbance deviation (i.e. +0.6% moisture content deviation). The unit solves the linear equations based on the Jacobian matrix and of the response manifold model at the current operating point to calculate the optimal mix adjustment (e.g. decrease 0.1% oil stone ratio, increase 0.1596 filler) required to offset the +0.6% moisture effect.

[0099] Before the correction is executed, the module performs a validation step: it calls the real-time performance evaluation unit to predict the performance under and conditions and confirm that all the constraints in the performance requirement package are still satisfied. After the validation is passed, the correction instruction is sent to the central control system of the mixing plant, which adjusts the speed of the asphalt pump and the discharge rate of the filler bin to achieve real-time dynamic correction of the production mix.

Claims

1. A system for intelligent calculation and scheme generation of road construction material proportions, characterized in that, include: The prior knowledge transfer module is used to retrieve and select one or more historical models from a historical manifold model library based on the input new project metadata, and output the parameters of the selected models as prior model parameters. The response manifold construction module under multi-scale constraints receives prior model parameters and constructs a multi-dimensional response manifold model that characterizes the mapping relationship between material composition, operating condition perturbation and macroscopic performance. The reverse generation module for the mix proportion scheme receives the multidimensional response manifold model and the performance requirement package input by the external user. The performance requirement package contains constraints on material properties. The module performs reverse solving to calculate the benchmark mix proportion scheme that satisfies the performance constraints in the performance requirement package. The active learning and model iteration module analyzes the uncertainty of the multidimensional response manifold model and calculates the parameter point with the largest model prediction variance. The parameter point is recommended for subsequent physical experiments, and feedback is used to establish new data points and iteratively update the multidimensional response manifold model. The construction process adaptive correction module, based on the baseline mix design output by the reverse generation module of the mix design, receives real-time monitored construction site disturbance parameters, and dynamically adjusts the material mix ratio according to actual measurement data and a multidimensional response manifold model. The construction process adaptive correction module includes: The data acquisition unit is used to monitor and collect key disturbance parameters at the construction site in real time. The real-time performance evaluation unit outputs the predicted material properties under the current perturbation conditions based on the trained multidimensional response manifold model. The correction decision unit calculates the amount of material ratio that needs to be adjusted based on the error between the current predicted material performance value and the performance constraints in the performance requirement package, and generates an optimized material ratio. Before outputting the optimized material ratio, the correction decision unit further performs a verification step, which includes: taking the optimized material ratio as input, calling the real-time performance evaluation unit to predict the expected performance under the optimized ratio, and determining whether the expected performance meets the performance constraints in the performance requirement package.

2. The intelligent calculation and scheme generation system for road construction material proportioning according to claim 1, characterized in that, The prior knowledge transfer module includes The historical manifold model library unit is used to store the response manifold model parameters and metadata of historical engineering projects. The similarity assessment and model selection unit is used to calculate the similarity between the metadata of the new project and the metadata of the historical projects, thereby selecting the most relevant historical model.

3. The intelligent calculation and scheme generation system for road construction material proportioning according to claim 1, characterized in that, The response manifold construction module under multi-scale constraints includes: The parameter space definition unit is used to define the value range of the proportioning parameter and the disturbance parameter; Microstructure constraint units are used to analyze the microscopic behavior of materials through physical model analysis tools to ensure the physical rationality of generating the multidimensional response manifold model; The manifold modeling and training unit is used to construct a Gaussian process regression model and train the multidimensional response manifold model based on known data points.

4. The intelligent calculation and scheme generation system for road construction material proportioning according to claim 1, characterized in that, The reverse generation module for the proportioning scheme includes: The requirement definition unit is used to receive and parse the performance requirement package and extract the performance constraints from the performance requirement package; The feasible region location unit is used to search for a set of feasible parameters that satisfy the performance constraints in the performance requirement package within the ratio parameter space; A robust optimization unit is used to seek an optimal ratio scheme based on the set of feasible parameters.

5. The intelligent calculation and scheme generation system for road construction material proportioning according to claim 1, characterized in that, The active learning and model iteration module includes: An uncertainty quantification unit is used to evaluate the prediction uncertainty of the output of the multidimensional response manifold model; The test site recommendation unit is used to determine the next most informative physical test site; The data feedback and model update unit is used to receive new experimental data and update the multidimensional response manifold model.

6. The intelligent calculation and scheme generation system for road construction material proportioning according to claim 1, characterized in that, The data acquisition unit is used to monitor multiple key disturbance parameters at the construction site in real time, including but not limited to temperature, humidity, and the physical properties of construction materials.

Images