Road surface performance prediction method and device and electronic equipment

Abstract

The invention discloses a pavement performance prediction method and device and electronic equipment, and relates to the technical field of industrial big data, and the method comprises the steps: carrying out the preprocessing of historical pavement data, and constructing a multi-level feature system which comprises key original features; training a plurality of heterogeneous base learners based on the multi-level feature system, and generating element feature vectors through cross validation; splicing the element feature vector with the key original feature to obtain an enhanced feature vector; and inputting the enhanced feature vector into a target element learner, reconstructing a prediction space through an attention mechanism, and generating a pavement performance prediction value. And by combining a plurality of heterogeneous base learners, the problems of cognitive barriers and insufficient generalization of a single model are solved, so that the model can learn data laws from multiple angles. Prediction space reconstruction is realized through an attention mechanism, the crossing from single model limitation to dynamic fusion is realized, the prediction precision and stability are remarkably improved, and a scientific basis is provided for a pavement maintenance decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial big data, and particularly relates to a road surface performance prediction method and device and an electronic device. BACKGROUND

[0002] With the continuous progress of industrial data management technology, road surface performance prediction has entered a new era of data-driven. At present, the scale of highway network continues to expand, and the continuous increase of traffic load leads to accelerated degradation of road surface performance. The contradiction between limited maintenance resources and growing performance demand further highlights the importance of accurate prediction. Therefore, how to efficiently process road surface data is particularly crucial.

[0003] At present, the traditional single machine learning model has the inherent defect of insufficient generalization ability, the tree model is difficult to handle complex nonlinear relationship, the support vector machine is sensitive to parameter configuration, and the neural network needs a large amount of data support and is insufficient in stability. In the process of dealing with the road surface degradation influenced by multiple factors coupling, the single model is difficult to fully grasp the internal law. Different algorithms have their own characteristics in data understanding level, but lack effective collaborative integration mechanism and do not fully mine the data features, which leads to insufficient accuracy of road surface performance prediction.

[0004] Therefore, how to improve the accuracy of road surface performance prediction is a problem to be solved at present. SUMMARY

[0005] The main purpose of the present application is to provide a road surface performance prediction method and device and an electronic device, aiming at solving the technical problem of insufficient accuracy of road surface performance prediction.

[0006] To achieve the above purpose, the present application provides a road surface performance prediction method, which comprises: preprocessing historical road surface data, constructing a multi-level feature system, the multi-level feature system comprising key original features; training a plurality of heterogeneous base learners based on the multi-level feature system, and generating an element feature vector through cross-validation; splicing the element feature vector and the key original features to obtain an enhanced feature vector; inputting the enhanced feature vector into a target meta-learner, reconstructing the prediction space through an attention mechanism, and generating a road surface performance prediction value.

[0007] In an embodiment, the step of preprocessing historical road surface data and constructing a multi-level feature system comprises: obtaining basic features, the basic features comprising road surface condition data, road attribute data, traffic load data and climate environment data; Based on the aforementioned basic features, multiple derived features are constructed. These derived features are used to characterize the physical or chemical processes of pavement performance degradation. The derived features include interaction features, scaling features, and transformation features. Based on the basic features and the derived features, a multi-level feature system is constructed.

[0008] In one embodiment, the heterogeneous base learner includes a first base learner, a second base learner, and a third base learner. The step of training multiple heterogeneous base learners based on the multi-level feature system and generating meta-feature vectors through cross-validation includes: K-fold cross-validation is performed on the first base learner, the second base learner and the third base learner respectively to obtain the corresponding out-of-fold prediction value sequence, which includes the first out-of-fold prediction value sequence, the second out-of-fold prediction value sequence and the third out-of-fold prediction value sequence. The first, second, and third out-of-division predicted value sequences are combined to obtain the meta-feature vector.

[0009] In one embodiment, the step of performing K-fold cross-validation on the first base learner, the second base learner, and the third base learner respectively to obtain the corresponding out-of-fold prediction sequence includes: The multi-level feature system is input into the first base learner, and the first predicted value is obtained through probability prediction, thus obtaining the first out-of-bounds predicted value sequence. The multi-level feature system is input into the second base learner, and the second predicted value is obtained through the attention mechanism, resulting in the second out-of-fold prediction value sequence. The multi-level feature system is input into the third base learner, and the third predicted value is obtained through the gradient boosting algorithm, resulting in the third out-of-division predicted value sequence.

[0010] In one embodiment, the step of concatenating the meta-feature vector with the key original feature to obtain the enhanced feature vector includes: Based on feature importance assessment, the top M key original features in terms of importance are selected from the multi-level feature system, where M is a preset integer; The key original features are concatenated with the meta-feature vector to form the enhanced feature vector.

[0011] In one embodiment, the step of inputting the enhanced feature vector into the target meta-learner and reconstructing the prediction space through an attention mechanism to generate a road performance prediction value includes: The hyperparameters of the initial meta-learner are optimized using the cuckoo search algorithm to obtain the target meta-learner; The enhanced feature vector is input into the target meta-learner, and the enhanced feature vector is embedded into the high-dimensional semantic space through a linear projection layer to obtain the projection sequence; The learnable position encoding matrix is ​​added to the projection sequence to obtain the embedding representation; The embedded representation is input into the Transformer encoder layer, and feature recalibration is performed through a multi-head self-attention mechanism to obtain the encoded sequence representation. The feature recalibration is used for context-awareness. The starting position feature vector is extracted from the encoded sequence representation to generate adaptive fusion weights for each base learner, wherein the starting position feature vector corresponds to the starting position of the sequence; The prediction results of each base learner are weighted and fused based on the adaptive fusion weights to generate road performance prediction values.

[0012] In one embodiment, the step of optimizing the hyperparameters of the initial meta-learner using the cuckoo search algorithm to obtain the target meta-learner includes: Initialize the search population and set the maximum number of iterations and the discovery probability. The search population includes multiple hyperparameter combinations. In each iteration, each combination of hyperparameters in the current population is evaluated, and the prediction error is calculated. The intermediate population is obtained by updating the hyperparameter combinations in the current population using the Levy flight strategy; Based on the intermediate population, some hyperparameter combinations are discarded according to the discovery probability, and new hyperparameter combinations are randomly generated to replace them, forming a new search population; Based on the new search population, return to the step of evaluating each hyperparameter combination in the current population until the maximum number of iterations is reached; The hyperparameter combination with the smallest prediction error is selected from the historical hyperparameter combinations as the target hyperparameter combination, and the target hyperparameter combination is configured in the target meta-learner.

[0013] In one embodiment, after inputting the enhanced feature vector into the target meta-learner, reconstructing the prediction space through an attention mechanism, and generating a road performance prediction value, the process includes: The predicted pavement performance values ​​are evaluated using multiple evaluation indicators, including the coefficient of determination, root mean square error, mean absolute error, and prediction interval coverage.

[0014] Furthermore, to achieve the above objectives, this application also proposes a pavement performance prediction device, which includes: The data processing module is used to preprocess historical road surface data and construct a multi-level feature system, which includes key original features. The vector generation module is used to train multiple heterogeneous base learners based on the multi-level feature system and generate meta-feature vectors through cross-validation. The vector enhancement module is used to concatenate the meta-feature vector with the key original feature to obtain the enhanced feature vector; The performance prediction module is used to input the enhanced feature vector into the target meta-learner, reconstruct the prediction space through an attention mechanism, and generate road performance prediction values.

[0015] In addition, to achieve the above objectives, this application also proposes an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the road performance prediction method as described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: Historical road surface data is preprocessed to construct a multi-level feature system, which includes key original features. By constructing a multi-level feature system, the problem of feature information silos in traditional methods is solved, enabling features to comprehensively capture the physical mechanisms and temporal cumulative effects of road surface decay.

[0017] This model trains multiple heterogeneous base learners based on a multi-level feature system and generates meta-feature vectors through cross-validation. By employing a combination of heterogeneous base learners, it overcomes the cognitive barriers and generalization limitations of single models, enabling the model to learn data patterns from multiple perspectives, including probabilistic prediction, feature selection, and anti-overfitting. The use of cross-validation to generate meta-feature vectors ensures unbiased prediction, resolves overfitting issues caused by data leakage, and ensures that meta-features reliably represent the model's collective cognition.

[0018] By concatenating the meta-feature vector with the key original features, an enhanced feature vector is obtained. The information density is enhanced through vector fusion, which solves the problem of missing contextual information in simple model fusion, enabling the input to simultaneously retain the key semantics of the original features and the abstract cognition of the meta-features.

[0019] The enhanced feature vector is input into the target meta-learner, and the prediction space is reconstructed through an attention mechanism to generate pavement performance predictions. By reconstructing the prediction space through the attention mechanism and dynamically calculating the fusion weights, the lack of adaptability in static weighted fusion is solved. This allows the model to intelligently adjust the contributions of each base learner based on the global context, achieving a leap from the limitations of a single model to dynamic intelligent fusion. This significantly improves prediction accuracy and robustness, providing a scientific basis for pavement maintenance decisions. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the first embodiment of the pavement performance prediction method of this application; Figure 2 This is a schematic diagram of the prediction results of the first base learner in the embodiments of this application; Figure 3 This is a schematic diagram of the prediction results of the second base learner in the embodiments of this application; Figure 4 This is a schematic diagram of the prediction results of the third base learner in the embodiments of this application; Figure 5 This is a schematic diagram of the prediction results of the heterogeneous model in the embodiments of this application; Figure 6 This is a schematic diagram illustrating the optimization and convergence process of the cuckoo search algorithm in an embodiment of this application; Figure 7 This is a comparison chart of the determination coefficients of various models in the embodiments of this application; Figure 8 This is a comparison chart of PICP indicators in the embodiments of this application; Figure 9 This is a schematic diagram of the road performance prediction device of this application.

[0023] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0025] This application provides a method for predicting road surface performance, referring to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the pavement performance prediction method of this application.

[0026] In this embodiment, the road performance prediction method includes steps S10 to S40: Step S10: Preprocess the historical road surface data to construct a multi-level feature system, which includes key original features.

[0027] It should be noted that historical road surface data refers to the raw dataset collected from the road management system. For example, preprocessing may include cleaning the raw dataset, constructing multi-level features based on the raw dataset, and then standardizing the multi-level features to obtain a multi-level feature system. Cleaning the raw dataset, such as imputing missing values ​​or correcting outliers, can be done by converting the data to a distribution with a mean of 0 and a standard deviation of 1, thus obtaining a multi-level feature system. Key raw features can be understood as selecting important feature subsets from the multi-level feature system.

[0028] Step S20: Train multiple heterogeneous base learners based on a multi-level feature system, and generate meta-feature vectors through cross-validation.

[0029] It should be noted that heterogeneous base learners refer to multiple heterogeneous machine learning models. Cross-validation can be used to ensure the unbiasedness of the data. Meta-feature vectors can be understood as vectors composed of multiple prediction results generated by multiple heterogeneous base learners through cross-validation.

[0030] Step S30: Concatenate the meta-feature vector with the key original features to obtain the enhanced feature vector.

[0031] It should be noted that the meta-feature vector is concatenated with the key original features to form a higher-dimensional fusion vector. The enhanced feature vector is a unified feature representation obtained after concatenation, containing both the metacognitive information of the base learner and retaining the key contextual information of the original features. This vector serves as the input to the target meta-learner, providing a rich semantic foundation for the reconstruction of the prediction space.

[0032] Step S40: The enhanced feature vector is input into the target meta-learner, and the prediction space is reconstructed through the attention mechanism to generate road performance prediction values.

[0033] It should be noted that the target meta-learner is a pre-trained meta-learner responsible for fusing the outputs of the base learners. The attention mechanism is used to globally perceive the dependencies between elements in the input sequence. Prediction space reconstruction can be understood as the meta-learner not operating in a fixed prediction space composed of the original features, but dynamically constructing a new, enhanced space that integrates the collective cognition of the base learners and key original contextual information, where the final decision is made. The collective cognition of the base learners is the meta-feature vector. The essence of prediction space reconstruction is actually a paradigm shift from static "feature-driven" prediction to dynamic decision-making driven by "cognition and context." The pavement performance prediction value refers to the final output of the current year's Pavement Condition Index (PCI) point prediction value, used to support maintenance decisions.

[0034] In this embodiment, by constructing a multi-level feature system, the problem of isolated feature information in traditional methods is solved, enabling features to comprehensively capture the physical mechanisms and temporal cumulative effects of pavement decay. The use of heterogeneous base learners overcomes the cognitive barriers and insufficient generalization of single models, allowing the model to learn data patterns from multiple perspectives, including probabilistic prediction, feature selection, and anti-overfitting. The use of cross-validation to generate meta-feature vectors ensures unbiased prediction, solves the overfitting problem caused by data leakage, and ensures that meta-features reliably represent the model's collective cognition. Furthermore, the meta-feature vectors are concatenated with key original features, and information density is enhanced through vector fusion, solving the problem of missing contextual information in simple model fusion. This allows the input to simultaneously retain the key semantics of the original features and the abstract cognition of the meta-features. An attention mechanism is used to reconstruct the prediction space and dynamically calculate the fusion weights, solving the problem of lack of adaptability in static weighted fusion. This allows the model to intelligently adjust the contributions of each base learner based on the global context, achieving a leap from the limitations of a single model to dynamic intelligent fusion, significantly improving prediction accuracy and robustness, and providing a scientific basis for pavement maintenance decisions.

[0035] In the second embodiment, step S10 includes: Step S101: Obtain basic features, including road surface condition data, road attribute data, traffic load data, and climate environment data.

[0036] It should be noted that basic characteristics refer to the original variables directly extracted from historical pavement data. For example, pavement condition data includes the previous year's Pavement Condition Index (PCI), used to reflect the historical performance status of the pavement. Road attribute data may include surface layer thickness, surface layer material, base layer thickness, base layer material, and pavement age, used to describe the changes in pavement structural properties over time. Traffic load data may include Annual Average Daily Traffic (AADT) and Annual Average Daily Truck Traffic (AADTT), used to quantify the cumulative impact of vehicle loads on the pavement. Climate environment data may include the number of high-temperature days, consecutive high-temperature days, low-temperature days, consecutive low-temperature days, and precipitation, used to capture the degradation effects of environmental stresses such as temperature and precipitation on pavement materials. Daily average temperatures >25℃ and <0℃ are defined as high and low temperatures, respectively, as detailed in Table 1.

[0037] Table 1 Step S102: Based on the basic features, construct multiple derived features. The derived features are used to characterize the physical or chemical processes of pavement performance degradation. The derived features include interaction features, proportional features, and transformation features.

[0038] It should be noted that derived features can be understood as new features generated through mathematical transformations or combinations of basic features, used to characterize the inherent laws governing pavement performance degradation. Interaction features can include traffic-condition interaction terms and climate-stress interaction terms, used to quantify the coupling effects of multiple factors. Proportional features can include traffic composition ratios, used to reflect the contribution of heavy vehicle proportions to pavement damage. Transformation features can include PCI nonlinear transformation terms and PCI decay rates, used to introduce nonlinear relationships or time-series dynamic characteristics. Furthermore, derived features can also include overall features, such as total pavement thickness.

[0039] For example, the specific mathematical definitions and physical meanings of the seven derived features can be expressed as follows: to .

[0040] in, This represents the traffic-condition interaction term, which characterizes the cumulative effect of traffic load on existing road conditions. This refers to the road surface condition index for the previous year. This represents the average daily traffic volume per year.

[0041] in, This is a climate stress interaction term used to quantify the combined stress on pavement materials caused by alternating high and low temperatures. This represents the total number of days with high temperatures throughout the year. This represents the total number of days with low temperatures throughout the year.

[0042] in, The traffic composition ratio reflects the proportion of heavy vehicles in traffic flow and is directly related to road surface damage. This represents the average daily truck traffic volume per year.

[0043] in, PCI decay rate is used to quantify the historical rate of decay of pavement performance. Indicates the road age.

[0044] in, The total thickness of the pavement represents the overall load-bearing capacity of the pavement structure. For the thickness of the base layer, This refers to the surface layer thickness.

[0045] in, Cumulative traffic intensity reflects the cumulative heavy traffic load borne by the road over its entire lifespan.

[0046] in, This is a PCI nonlinear transformation term used to capture potential nonlinear decay modes of road condition indicators.

[0047] Step S103: Construct a multi-level feature system based on basic features and derived features.

[0048] For example, 13 basic features and 7 derived features can be integrated into a hierarchical structure. By standardizing the data to unify the dimensions, the features can be made comparable in the model, providing a systematic input for subsequent model training.

[0049] In this embodiment, multiple sources of basic features, including pavement conditions, road attributes, traffic load, and climate environment, are acquired, solving the problems of single data dimension and isolated features in traditional methods, and achieving comprehensive coverage of factors affecting pavement performance. By constructing derived features such as interactive features, proportional features, and transformation features, and through feature engineering driven by physical meaning, the problem of insufficient characterization of nonlinear decay processes by original features is solved, enabling a deep analysis of the pavement performance evolution mechanism. By hierarchically integrating original and derived features and standardizing them, the problems of a chaotic and unsystematic feature system are solved, achieving the structuring of feature information, significantly enhancing the richness and physical interpretability of the feature space, providing a more comprehensive input foundation for subsequent models, and thus directly supporting the improvement of prediction accuracy.

[0050] In the third embodiment, the heterogeneous base learner includes a first base learner, a second base learner, and a third base learner. For example, the first base learner can be a probabilistic prediction model, such as Natural Gradient Boosting (NGBoost), which boosts the output probability distribution through natural gradients, supporting uncertainty quantification. The second base learner can be an attention mechanism model, such as TabNet, which achieves soft feature selection through sequential attention, balancing interpretability and nonlinear modeling. The third base learner can be a gradient boosting algorithm model, such as LGBM-DART (Light Gradient Boosting Machine-Dropouts meet Additive Regression Trees), which introduces the Dropout mechanism to enhance generalization ability and can efficiently process large-scale data. Dropout is a regularization technique used to mitigate overfitting. Based on this, step S20 includes:

[0051] Step S201: Perform K-fold cross-validation on the first base learner, the second base learner, and the third base learner to obtain the corresponding out-of-fold prediction sequence. The out-of-fold prediction sequence includes the first out-of-fold prediction sequence, the second out-of-fold prediction sequence, and the third out-of-fold prediction sequence.

[0052] Specifically, step S201 includes inputting the multi-level feature system into the first base learner, obtaining the first predicted value through probability prediction, and obtaining the first out-of-bounds predicted value sequence; inputting the multi-level feature system into the second base learner, obtaining the second predicted value through the attention mechanism, and obtaining the second out-of-bounds predicted value sequence; inputting the multi-level feature system into the third base learner, obtaining the third predicted value through the gradient boosting algorithm, and obtaining the third out-of-bounds predicted value sequence.

[0053] It should be noted that the multi-level feature system can be divided into training fold data and validation fold data. K-fold cross-validation is performed on the first base learner. Specifically, in each fold, the first base learner is trained using the training fold data, and the first base learner's prediction value for that validation fold is obtained by probability prediction using the validation fold data. The prediction values ​​of all folds are merged to form the sequence of prediction values ​​outside the first fold. K-fold cross-validation is performed on the second base learner. Specifically, in each fold, the second base learner is trained using the training fold data, and the second base learner's prediction value for the validation fold is obtained through an attention mechanism using the validation fold data. The prediction values ​​of all folds are merged to form a sequence of prediction values ​​outside the second fold. K-fold cross-validation is performed on the third base learner. Specifically, in each fold, the third base learner is trained using the training fold data, and the prediction value of the third base learner for that validation fold is obtained using the gradient boosting algorithm using the validation fold data. The prediction values ​​of all folds are merged to form the sequence of prediction values ​​outside the third fold.

[0054] It's important to note that K-fold cross-validation is a model validation technique that randomly divides the dataset into K mutually exclusive subsets. The outer loop is a K-iteration framework, with each iteration using K-1 subsets as the training set and one subset as the validation set. In each training iteration of the inner loop, the base learner learns independently on the training set and generates predictions on the validation set. This process ensures that each sample is used for validation only once, avoiding data leakage. The out-of-fold predicted value sequence refers to the sequence of prediction results generated by K-fold cross-validation. The first out-of-fold predicted value sequence can consist of the predictions output by NGBoost on the validation set in each outer loop, with a sequence length equal to the number of samples, and each prediction value corresponding to the mean of the probability distribution of the sample. The second out-of-fold predicted value sequence can be generated by TabNet under the attention mechanism, reflecting the output after feature soft selection. The third out-of-fold predicted value sequence can be generated by LGBM-DART using the gradient boosting algorithm, incorporating the Dropout mechanism to reduce overfitting.

[0055] For example, K can be 5, the number of trees in NGBoost can be 200, the learning rate can be set to 0.02, and the probability distribution can be output using natural gradients. Specifically, the prediction process of NGBoost can be described as outputting the parameters of a probability distribution. For the ... The sample has an output that is a Gaussian distribution with a mean. The corresponding formula is:

[0056] in, For the first The input feature vector of each sample, For the mapping function of the NGBoost model, For the first NGBoost predictions for each sample.

[0057] TabNet sets the decision layer width to 8, the attention layer width to 8, the γ coefficient to 1.3, the optimizer Adam, and the learning rate to 0.02. After feature soft selection via a sequential attention mechanism, the output point prediction value is calculated using the following formula: in, For the mapping function of the TabNet model, For the first TabNet predictions for each sample.

[0058] The LGBM-DART tree can have 300 trees, a learning rate of 0.02, a maximum tree depth of 5, 20 leaves, a dropout rate of 5%, and a skip-drop probability of 80%. A Dropout mechanism is introduced, and a multiple additive regression tree algorithm is used to enhance generalization ability. The output point prediction value is given by the following formula:

[0059] in, For the mapping function of the LGBM-DART model, For the first LGBM-DART predicted values ​​for each sample.

[0060] Step S202: Combine the first out-of-division predicted value sequence, the second out-of-division predicted value sequence, and the third out-of-division predicted value sequence to obtain the meta-feature vector.

[0061] For example, for each sample The prediction results of its three base learners are used to construct a 3-dimensional meta-feature vector. The corresponding formula is: in, For the first The meta-feature vector of each sample encapsulates the collective understanding of road performance from the three heterogeneous models and forms the basis for subsequent information fusion.

[0062] In this embodiment, K-fold cross-validation is used to process heterogeneous base learners. The outer loop framework divides the data subset, while the inner loop ensures that each base learner independently generates out-of-fold predictions, solving the overfitting problem caused by data leakage in traditional training and achieving unbiased estimation of the meta-feature vector. By quantifying uncertainty with the first base learner, implementing soft feature selection with the second base learner, and enhancing generalization ability with the third base learner, the limitations of single-model cognition and feature processing are overcome, enabling multi-view data learning and complementary advantages. Furthermore, by aligning and concatenating the three sequences according to samples to form the meta-feature vector, the problems of information isolation and bias accumulation in simple model fusion are solved, achieving reliable fusion of heterogeneous model cognition.

[0063] Based on the above embodiments and implementation methods, a fourth embodiment of the pavement performance prediction method of this application is proposed. In the fourth embodiment, step S30 includes:

[0064] Step S301: Based on feature importance assessment, select the top M key original features from the multi-level feature system, where M is a preset integer.

[0065] It should be noted that the Random Forest algorithm can be used to score the importance of all features in the multi-level feature system. The importance score ranking is generated by calculating the average contribution of each feature to the improvement of road performance when splitting nodes in the Random Forest decision tree. The key original features represent the top M most important features selected from the multi-level feature system based on the feature importance assessment results. The preset integer M refers to a hyperparameter pre-set according to the actual data scale and model complexity requirements, used to control the number of key features and balance information density with computational efficiency.

[0066] Step S302: Concatenate the key original features with the meta-feature vector to form an enhanced feature vector.

[0067] For example, the 3D meta-feature vector of each sample Selected from 20 original features Key original features They are pieced together to form a unified whole. 3D Enhanced Feature Vector Subsequently, this vector is considered as a vector of length... Input sequences with a feature dimension of 1 , is represented as:

[0068] in, For the first A subset of key original features of each sample For the first Enhanced feature vectors for each sample.

[0069] In this embodiment, based on the feature importance assessment of random forest, a data-driven objective scoring mechanism is used to screen key original features, solving the problems of strong subjectivity, feature redundancy, or information loss in traditional feature selection, and achieving efficient focusing on high-discriminative features. By presetting an integer M to balance feature dimensionality and model complexity, the problem of excessive computational burden or noise introduction caused by high-dimensional feature input is solved, achieving optimized compression of the feature space. Concatenating the meta-feature vector with the key original features achieves the structured integration of heterogeneous information, directly supporting the meta-learner's accurate reconstruction of the prediction space.

[0070] Based on the above embodiments and implementation methods, a fifth embodiment of the pavement performance prediction method of this application is proposed. In the fifth embodiment, step S40 includes:

[0071] Step S401: Optimize the hyperparameters of the initial meta-learner using the Cuckoo Search algorithm to obtain the target meta-learner.

[0072] It should be noted that the Cuckoo Search algorithm can generate new solutions randomly and update the search population probabilistically for efficient global search. In this embodiment, it optimizes the hyperparameters of the initial meta-learner by minimizing the root mean square error (RMSE) of the validation set. The initial meta-learner can be understood as an unoptimized Transformer architecture model, whose hyperparameters can be set by default or randomly, serving as the starting point for optimization. Hyperparameters refer to the structural parameters of the meta-learner, including the hidden layer dimension, the number of attention heads, the number of encoder layers, the feedforward network dimension, etc. The target meta-learner can be a Transformer meta-learner.

[0073] Step S402: Input the enhanced feature vector into the target meta-learner, and embed the enhanced feature vector into the high-dimensional semantic space through a linear projection layer to obtain the projection sequence.

[0074] For example, a sequence can be projected using a linear projection layer. Embedded into a high-dimensional semantic space ,in This represents the dimension of the model's hidden layers.

[0075] Step S403: Add the learnable position encoding matrix to the projection sequence to obtain the embedding representation.

[0076] For example, a learnable position encoding matrix is ​​introduced. And add it to the embedded sequence, with the corresponding formula as follows: in, For the reason The input sequence constituted This is the embedding layer weight matrix. For the position encoding matrix, This is the sequence after embedding and adding positional encoding. It should be noted that... It is the projection weight matrix. Position encoding. This allows the Transformer to distinguish features at different positions in a sequence. For example, the first position is the NGBoost output, and the second position is the... Each location is a key original feature, and we understand the structural relationships between them.

[0077] Step S404: The embedded representation is input into the Transformer encoder layer, and the feature is recalibrated through a multi-head self-attention mechanism to obtain the encoded sequence representation. The feature recalibration is used for context-awareness.

[0078] For example, will Input a by A Transformer module consisting of stacked layer encoders. Each layer encoder contains a multi-head self-attention mechanism and a feedforward neural network. For the ... Layer, Input Transformed linearly into a query matrix Key matrix Sum matrix , is represented as:

[0079] The calculation for each attention head is as follows: in, For the first The output of the layer encoder, These are weight matrices for the query, key, and value, respectively. Let be the dimension of the key vector. These are the query, key, and value matrices, respectively. All The outputs of the attention heads are concatenated, linearly projected, and then passed to the feedforward network, as shown below:

[0080] in, For the first The output of each attention head, For the first The output of the self-attention mechanism in the layer encoder, This is the linear projection weight matrix used to fuse the outputs of multiple attention heads.

[0081] It should be noted that the feedforward network consists of two linear transformations and a ReLU activation function in between, used to perform nonlinear transformations of features.

[0082] in, For the first The output of the layer self-attention mechanism, For the weights and biases of the feedforward network, For the first The final output of the layer encoder. Through dynamically calculated weights using a self-attention mechanism, the model can evaluate the dependencies between all input information within the global context.

[0083] Step S405: Extract the start position feature vector from the encoded sequence representation to generate adaptive fusion weights for each base learner. The start position feature vector corresponds to the start position of the sequence.

[0084] For example, after After layer Transformer encoding, a sequence representation rich in contextual information can be obtained. You can take The feature vector corresponding to a special marker [CLS] pre-added at the beginning of the sequence. This vector is considered to aggregate global information from the entire sequence. Finally, it is passed through an output layer consisting of a fully connected layer and a softmax function to generate fusion weights for the three base learners. .

[0085] Step S406: The prediction results of each base learner are weighted and fused based on adaptive fusion weights to generate road performance prediction values.

[0086] For example, the final PCI prediction value is obtained by a weighted sum of the prediction results of the base learners, and the corresponding formula is: in, The feature vector corresponding to the [CLS] label, which aggregates global sequence information. The adaptive fusion weight vector of the base learners This is the final PCI prediction. The prediction results of a single base learner versus the pre-ordered model are referenced. Figures 2 to 5 ,in Figure 2 Corresponding to the first base learner, Figure 3 Corresponding to the second base learner, Figure 4 Corresponding to the third basis learner, Figure 5This corresponds to a heterogeneous model. A heterogeneous model is a comprehensive model that combines three base learners and a target meta-learner.

[0087] In this embodiment, the Cuckoo Search algorithm is used to optimize the hyperparameters of the meta-learner, and the enhanced feature vectors are embedded into a high-dimensional semantic space through a linear projection layer, solving the problem of insufficient expressive power of low-dimensional features and realizing rich representation of feature semantics. The learnable position encoding matrix is ​​added to the projection sequence, and positional information is introduced to distinguish the feature order, solving the problem of missing structural relationships in serialized input and realizing context-aware embedding representation. By employing a Transformer encoder layer and a multi-head self-attention mechanism for feature recalibration, and through dynamic weight calculation in multiple subspaces, the problem of context ignoring in static fusion is solved, realizing intelligent weighting of features based on global dependencies. By employing a technique of extracting the feature vector of the starting position to generate adaptive fusion weights, and by aggregating global information of the sequence to generate dynamic weights, the problem of lack of flexibility in fixed-weight fusion is solved, realizing context-adaptive integration of base learner predictions. By employing a weighted fusion technique to generate road performance prediction values, and through linear combination of weights and prediction values, the accuracy limitations of simple averaging or voting fusion are solved, achieving accurate optimization of prediction results. Through prediction space reconstruction and context-aware fusion, the accuracy and robustness of road performance prediction are significantly improved.

[0088] In one implementation, step S401 includes: initializing a search population and setting a maximum number of iterations and a discovery probability, the search population including multiple hyperparameter combinations; in each iteration, evaluating each hyperparameter combination in the current population and calculating the prediction error; updating the hyperparameter combinations in the current population using the Lévy flight strategy to obtain an intermediate population; based on the intermediate population, discarding some hyperparameter combinations according to the discovery probability and randomly generating new hyperparameter combinations to replace them, forming a new search population; based on the new search population, returning to the step of evaluating each hyperparameter combination in the current population until the maximum number of iterations is reached; selecting the hyperparameter combination with the smallest prediction error from historical hyperparameter combinations as the target hyperparameter combination, and configuring the target hyperparameter combination in the target meta-learner.

[0089] It should be noted that the performance of the Transformer meta-learner is highly dependent on its hyperparameter configuration. The cuckoo search algorithm is used in the pre-defined search space. Find the optimal match For example, the population size is set to 15, the maximum number of iterations is 15, and the discovery probability is... The initial population is generated randomly. In each generation, the algorithm performs the following steps:

[0090] a. For each hyperparameter combination in the population For rapid evaluation, you can use Configure the meta-learner, perform training on a small number of batches, such as 3-5 batches, and compute its RMSE loss on a randomly sampled validation subset. You can take 100 samples.

[0091] b. Record and update the global best fitness and best solution. The optimal fitness is the lowest RMSE.

[0092] c. Generate new solutions using the Lévy flight strategy. The step size of the Lévy flight is determined by the following formula:

[0093] d. in terms of probability Discard some of the worse solutions and generate new random solutions to replace them.

[0094] in, and Let be a random variable that follows a standard normal distribution. For the Levy flight index parameters, .

[0095] For example, such as Figure 6 As shown, after 15 iterations, the Cuckoo optimization algorithm successfully converged. The initial best fitness (RMSE) was 0.318154, and the final best fitness improved to 0.300968. The algorithm outputs the globally optimal hyperparameter combination of the Transformer meta-learner as follows: hidden layer dimension (d_model): 376; number of attention heads (nhead): 8; number of encoder layers (num_layers): 2; feedforward network dimension (dim_feedforward): 752; dropout rate: 0.185; learning rate (learning_rate): 0.0002.

[0096] In this implementation, by randomly generating diverse hyperparameter combinations as a starting point and controlling the optimization cycle, the problems of blind manual parameter tuning and wasted computational resources are solved, achieving efficient startup and controllability of the search process. An intermediate population is obtained by updating the hyperparameter combinations through a Lévy flight strategy, and global exploration is achieved through a random step size of the Lévy distribution, overcoming the limitation of gradient-based methods being prone to local optima and achieving broad coverage of the search space. Simultaneously, based on the discovery probability, some hyperparameter combinations are discarded and new combinations are randomly generated to form a new population. By probabilistically eliminating inferior solutions and introducing randomness, the problems of premature convergence or loss of diversity in the population are solved, achieving a dynamic balance between global and local search. Iterative optimization solves the problem of insufficient optimization in a single run, achieving progressive refinement of hyperparameter configuration.

[0097] In one implementation, after step S40, the method includes: evaluating the predicted pavement performance value using multiple evaluation indicators, including the coefficient of determination, root mean square error, mean absolute error, and prediction interval coverage.

[0098] For example, the coefficient of determination (R²) is used to quantify the extent to which the model explains the variability of the target variable, and its formula is as follows: R²= in, For the sample size, For the true value, For predicted values, This is the mean of the true values.

[0099] The root mean square error (RMSE) measures the deviation between the model's predicted values ​​and the actual values. Its formula is as follows: RMSE= in, For the sample size, Let i be the true road surface condition index for the i-th sample. For the model to the first The predicted value for each sample.

[0100] The mean absolute error (MAE) measures the absolute value of prediction error and is insensitive to outliers. The formula is as follows: MAE= in, For the sample size, For the true value, These are predicted values.

[0101] Predicted Interval Coverage (PICP) is used to assess the reliability of a predicted interval, and its calculation formula is as follows: PICP= Where N is the number of test samples, and 1 is the indicator function. and These are the lower and upper bounds of the prediction interval generated by the model for the i-th sample. This is the true observation value of the i-th sample. It is set to 1 if the true value yi of the i-th sample is within its corresponding prediction interval, and 0 otherwise.

[0102] This implementation uses a pavement performance prediction method based on heterogeneous model cascade and prediction space reconstruction using an independent test set to evaluate performance. It employs a comprehensive comparative analysis of RMSE, MAE, R², and PICP indices, taking R² and PICP indices as examples. Figures 7 to 8It can be seen that, compared with single base learners such as NGBoost, TabNet, and LGBM-DART, the cascaded learning heterogeneous feature splicing model performs better on all evaluation metrics, proving that it can collaboratively optimize multivariate prediction modes through meta-learners, thereby achieving higher accuracy and stability in road condition index prediction.

[0103] Based on the same inventive concept, such as Figure 9 As shown, this application also provides a pavement performance prediction device, which includes: Data processing module 10 is used to preprocess historical road surface data and construct a multi-level feature system, which includes key original features. Vector generation module 20 is used to train multiple heterogeneous base learners based on the multi-level feature system and generate meta-feature vectors through cross-validation. Vector enhancement module 30 is used to concatenate the meta-feature vector with the key original feature to obtain an enhanced feature vector; The performance prediction module 40 is used to input the enhanced feature vector into the target meta-learner, reconstruct the prediction space through an attention mechanism, and generate road performance prediction values.

[0104] It should be noted that the pavement performance prediction device and the pavement performance prediction method provided in this application are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned pavement performance prediction method, and the repeated parts will not be described again.

[0105] In some embodiments, an electronic device provided in this application includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the above-described road performance prediction method.

[0106] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0107] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0108] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting pavement performance, characterized in that, The method includes: Historical road surface data is preprocessed to construct a multi-level feature system, which includes key original features; Multiple heterogeneous base learners are trained based on the multi-level feature system, and meta-feature vectors are generated through cross-validation. The meta-feature vector is concatenated with the key original feature to obtain the enhanced feature vector; The enhanced feature vector is input into the target meta-learner, and the prediction space is reconstructed through the attention mechanism to generate road performance prediction values.

2. The method as described in claim 1, characterized in that, The steps of preprocessing historical road surface data and constructing a multi-level feature system include: Acquire basic features, including road surface condition data, road attribute data, traffic load data, and climate environment data; Based on the aforementioned basic features, multiple derived features are constructed. These derived features are used to characterize the physical or chemical processes of pavement performance degradation. The derived features include interaction features, scaling features, and transformation features. Based on the basic features and the derived features, a multi-level feature system is constructed.

3. The method as described in claim 2, characterized in that, The heterogeneous base learner includes a first base learner, a second base learner, and a third base learner. The step of training multiple heterogeneous base learners based on the multi-level feature system and generating meta-feature vectors through cross-validation includes: K-fold cross-validation is performed on the first base learner, the second base learner and the third base learner respectively to obtain the corresponding out-of-fold prediction value sequence, which includes the first out-of-fold prediction value sequence, the second out-of-fold prediction value sequence and the third out-of-fold prediction value sequence. The first, second, and third out-of-bounds predicted value sequences are combined to obtain the meta-feature vector.

4. The method as described in claim 3, characterized in that, The step of performing K-fold cross-validation on the first base learner, the second base learner, and the third base learner respectively to obtain the corresponding out-of-fold prediction sequence includes: The multi-level feature system is input into the first base learner, and the first predicted value is obtained through probability prediction, thus obtaining the first out-of-bounds predicted value sequence. The multi-level feature system is input into the second base learner, and the second predicted value is obtained through the attention mechanism, resulting in the second out-of-fold prediction value sequence. The multi-level feature system is input into the third base learner, and the third predicted value is obtained through the gradient boosting algorithm, resulting in the third out-of-division predicted value sequence.

5. The method as described in claim 1, characterized in that, The step of concatenating the meta-feature vector with the key original features to obtain the enhanced feature vector includes: Based on feature importance assessment, the top M key original features in terms of importance are selected from the multi-level feature system, where M is a preset integer; The key original features are concatenated with the meta-feature vector to form the enhanced feature vector.

6. The method as described in claim 1, characterized in that, The step of inputting the enhanced feature vector into the target meta-learner and reconstructing the prediction space through an attention mechanism to generate road performance prediction values ​​includes: The hyperparameters of the initial meta-learner are optimized using the cuckoo search algorithm to obtain the target meta-learner; The enhanced feature vector is input into the target meta-learner, and the enhanced feature vector is embedded into the high-dimensional semantic space through a linear projection layer to obtain the projection sequence; The learnable position encoding matrix is ​​added to the projection sequence to obtain the embedding representation; The embedded representation is input into the Transformer encoder layer, and feature recalibration is performed through a multi-head self-attention mechanism to obtain the encoded sequence representation. The feature recalibration is used for context-awareness. The starting position feature vector is extracted from the encoded sequence representation to generate adaptive fusion weights for each base learner, wherein the starting position feature vector corresponds to the starting position of the sequence; The prediction results of each base learner are weighted and fused based on the adaptive fusion weights to generate road performance prediction values.

7. The method as described in claim 1, characterized in that, The step of optimizing the hyperparameters of the initial meta-learner using the cuckoo search algorithm to obtain the target meta-learner includes: Initialize the search population and set the maximum number of iterations and the discovery probability. The search population includes multiple hyperparameter combinations. In each iteration, each combination of hyperparameters in the current population is evaluated, and the prediction error is calculated. The intermediate population is obtained by updating the hyperparameter combinations in the current population using the Levy flight strategy; Based on the intermediate population, some hyperparameter combinations are discarded according to the discovery probability, and new hyperparameter combinations are randomly generated to replace them, forming a new search population; Based on the new search population, return to the step of evaluating each hyperparameter combination in the current population until the maximum number of iterations is reached; The hyperparameter combination with the smallest prediction error is selected from the historical hyperparameter combinations as the target hyperparameter combination, and the target hyperparameter combination is configured in the target meta-learner.

8. The method as described in claim 1, characterized in that, The step of inputting the enhanced feature vector into the target meta-learner, reconstructing the prediction space through an attention mechanism, and generating road performance prediction values ​​includes: The predicted pavement performance values ​​are evaluated using multiple evaluation indicators, including the coefficient of determination, root mean square error, mean absolute error, and prediction interval coverage.

9. A road surface performance prediction device, characterized in that, The road surface performance prediction device includes: The data processing module is used to preprocess historical road surface data and construct a multi-level feature system, which includes key original features. The vector generation module is used to train multiple heterogeneous base learners based on the multi-level feature system and generate meta-feature vectors through cross-validation. The vector enhancement module is used to concatenate the meta-feature vector with the key original feature to obtain the enhanced feature vector; The performance prediction module is used to input the enhanced feature vector into the target meta-learner, reconstruct the prediction space through an attention mechanism, and generate road performance prediction values.

10. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the road performance prediction method as described in any one of claims 1 to 8.

Images