A road surface condition degradation prediction method fusing road age perception and physical constraints

Abstract

The present application belongs to the technical field of road surface condition data prediction, and particularly relates to a road surface condition degradation prediction method fusing road age perception and physical constraints, collecting road surface management data of an un-maintained road section, and performing multi-source feature preprocessing on the road surface management data to obtain a road surface feature vector; introducing a physical information constraint module on the basis of a Transformer model, the physical information constraint module introducing physical laws for the Transformer model through a road surface condition degradation equation and interval constraints, thereby constructing and training a physical information Transformer model for road surface condition degradation prediction; and introducing a SHAP explainability framework to perform explainability analysis on the optimal physical information Transformer model. The present application can ensure that road surface condition prediction conforms to physical laws and has clear explainability.

Description

Technical Field

[0001] This invention relates to the field of pavement condition data prediction technology. Specifically, it relates to a pavement condition degradation prediction method that integrates pavement age perception and physical constraints. This method is applicable to long-term condition prediction of various types of pavements, such as highways and urban roads, and provides accurate data support for pavement maintenance decisions. Background Technology

[0002] The deterioration pattern of road surface conditions directly determines the timing of road maintenance, resource allocation efficiency, and service life. With the rapid increase in highway mileage, traditional reactive maintenance models can no longer meet the demands for low-cost, high-efficiency management. There is an urgent need for accurate condition deterioration prediction to anticipate declining road surface conditions and implement maintenance measures before conditions deteriorate to critical levels, thus avoiding high repair costs due to premature road damage. Currently, the mainstream road surface condition prediction methods are mainly divided into three categories:

[0003] Traditional physical models, which establish mathematical equations based on road surface decay mechanisms, can reflect physical laws, but rely on only a small number of structural and environmental parameters. They cannot capture the complex nonlinear relationships between multiple sources of features such as traffic load and multiple environmental factors, and their prediction accuracy is limited by the parameter fitting ability.

[0004] Statistical learning methods, such as linear regression and random forest, can handle multiple feature inputs, but they are weak in modeling temporal correlations and are difficult to adapt to long-term prediction scenarios.

[0005] Deep learning methods, such as LSTM (Long Short-Term Memory) networks and ordinary Transformer models, can model the temporal dependence and global correlation of multi-source features, but they only optimize the model by fitting data and ignore the inherent physical laws of road surface condition decay. This may lead to prediction results that exceed the reasonable range of engineering experience, and the model has poor interpretability and cannot provide quantitative basis for feature contribution for maintenance decisions.

[0006] In summary, existing methods cannot simultaneously meet the requirements of multi-source feature modeling capability, consistency of physical laws, rationality of engineering applications, and interpretability. There is an urgent need for a road condition prediction technology that balances data-driven accuracy with physical mechanism constraints. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting road surface condition degradation that integrates road age perception and physical constraints, in order to solve the problems mentioned in the background art.

[0008] To achieve the above technical objectives, the present invention adopts the following technical solution: a method for predicting road surface condition degradation by integrating road age perception and physical constraints, comprising the following steps:

[0009] S1. Collect pavement management data of unmaintained road sections and perform multi-source feature preprocessing on the pavement management data to obtain pavement feature vectors. All pavement management data are processed to obtain different pavement feature vectors to form a dataset.

[0010] S2. Based on the Transformer model, a physical information constraint module is introduced. The physical information constraint module introduces physical laws into the Transformer model through the road condition decay equation and interval constraints, thereby constructing a physical information Transformer model. The dataset is divided into a training set and a test set. The physical information Transformer model is trained using the training set to obtain the optimal hyperparameters. The test set is used for testing and verification. If the verification is successful, the optimal physical information Transformer model for road condition decay prediction is obtained.

[0011] S3. The SHAP interpretability framework is introduced to perform interpretability analysis on the optimal physical information Transformer model.

[0012] Further preferred, the multi-source feature preprocessing method is as follows: standardizing the continuous variables in the multi-source features; performing one-hot encoding on the categorical variables in the multi-source features; and performing separate standardization on the road age features.

[0013] Further preferably, the physical information Transformer model includes:

[0014] Input projection layer: Employs a fully connected network to map the road feature vectors obtained from multi-source feature preprocessing to the Transformer model dimension. , thus obtaining the mapped road surface feature vector;

[0015] Physically guided road age location coding layer: Road age sensing location coding is performed to obtain road age location coding vector. The road age location coding vector is then fused with the mapped road surface feature vector to obtain a road surface feature vector with fused road age information.

[0016] Transformer encoder: It consists of two stacked encoder layers. Each encoder layer contains a multi-head self-attention network and a feedforward network. After processing by the multi-head self-attention network, residual connections and normalization are performed. After processing by the feedforward network, residual connections and normalization are also performed before outputting to the next stage.

[0017] Output layer: The features extracted by the Transformer encoder are mapped to the normalized road condition index (PCI) prediction value, and then denormalized to restore the original scale prediction value, while introducing physical information constraints;

[0018] Physical Information Constraint Module: The physical information constraint module introduces physical laws into the Transformer model through road condition attenuation equations and interval constraints.

[0019] Further optimization, the physical guidance road age location coding layer processing includes:

[0020] First, road age is destandardized, and the original road age is restored based on the standardized road age.

[0021] Encoding vector generation: Based on the original road age, a sine-cosine function is used to generate the road age location encoding vector.

[0022] Further optimization is achieved through a composite loss function. The learnable parameters related to physical constraints in the Transformer model are collaboratively optimized, ensuring that the Transformer model both fits the data and follows physical laws. This is achieved through a composite loss function. Loss due to data fitting Physical law constraint loss Deviation upper limit penalty loss The composition, and the formula are:

[0023] ;

[0024] in, The loss weights are constrained by physical laws; The upper limit of the deviation penalty loss weight is determined by the interval constraint; Calculated based on the mean squared error between the predicted and actual values ​​from the Transformer model; Calculated based on the mean square error between the predicted and actual values ​​of the road surface condition degradation equation.

[0025] Further preferably, the predicted value of the road surface condition attenuation equation is calculated based on the following formula:

[0026] ;

[0027] in, This refers to the road surface condition index. This is the initial value of the pavement condition index, i.e., the pavement condition index value at t=0, where t is the pavement age; For learnable parameters related to the magnitude of initial decay of road surface conditions, For learnable parameters related to the rate of road surface condition decay, These are learnable parameters related to the stage characteristics of pavement decay.

[0028] Further preferred, the upper limit of the deviation penalty loss The calculation formula is:

[0029] ;

[0030] Where N is the number of samples, For the first The original scale pavement condition index prediction value of each sample is obtained by inverse standardization of the pavement condition index standardized value output by the physical information Transformer model; For the first Each sample is a predicted value based on the road condition attenuation equation.

[0031] Further preferred, the multi-source features include four categories: road surface condition features, traffic load parameters, environmental impact factors, and structural attributes, specifically:

[0032] Road surface condition characteristics: Road surface condition index (PCI) value of the previous year; Traffic load parameters: Average daily freight traffic volume (AADTT) and average daily traffic volume (AADT) of the year; Environmental impact factors: Total number of high temperature days, consecutive high temperature days, total number of low temperature days, consecutive low temperature days, and rainfall; Structural attributes: Base course thickness, base course material type, surface course thickness, surface course material type, and road age.

[0033] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the road condition degradation prediction method that integrates road age perception and physical constraints.

[0034] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the road condition degradation prediction method that integrates road age perception and physical constraints.

[0035] This invention, by fusing multi-source features with temporal priors on road age and differing from the general encoding of data sequence positions in ordinary Transformers, encodes road age based on the early slow decay and mid-term rapid decay patterns of road surface conditions, achieving the fusion of temporal priors and feature information. It constructs a physical information Transformer model and achieves collaborative parameter optimization, overcoming the core shortcomings of existing methods. Validated on 1544 road surface samples from expressways in Shanxi Province: prediction accuracy is significantly improved; interpretability is clear: SHAP analysis identifies road age and the previous year's PCI as core influencing factors, quantifying feature contributions; broad engineering applicability: it can be adapted to asphalt, cement pavement, and data from different climate regions, contributing to more scientific maintenance decisions and more rational allocation of funds. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a structural diagram of the Transformer model for physical information.

[0038] Figure 3 A comparison chart of ANN model predictions and actual values;

[0039] Figure 4 A graph showing the comparison between the predicted values ​​and the actual values ​​from the Transformer model;

[0040] Figure 5 A comparison chart of the predicted and actual values ​​from the Transformer model for physical information;

[0041] Figure 6 A comparison chart of the coefficients of determination;

[0042] Figure 7 This is a comparison chart of mean square error and mean absolute error;

[0043] Figure 8 This is a schematic diagram of Transformer interpretability analysis;

[0044] Figure 9 This is a schematic diagram of the interpretability analysis of physical information using Transformer. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] like Figure 1 As shown, the present invention provides a method for predicting road surface condition degradation by integrating road age perception and physical constraints, comprising the following steps;

[0047] S1. Data preparation: Collect pavement management data of unmaintained road sections and perform multi-source feature preprocessing on the pavement management data to obtain pavement feature vectors. All pavement management data are processed to obtain different pavement feature vectors to form a dataset.

[0048] S2. Model Building: A physical information constraint module is introduced based on the Transformer model. The physical information constraint module introduces physical laws into the Transformer model through the road condition decay equation and interval constraints, thereby constructing a physical information Transformer model. The dataset is divided into a training set and a test set. The physical information Transformer model is trained using the training set to obtain the optimal hyperparameters. The test set is used for testing and verification. If the verification is successful, the optimal physical information Transformer model for road condition decay prediction is obtained.

[0049] S3. Model Interpretation: The SHAP interpretability framework is introduced to perform interpretability analysis on the optimal physical information Transformer model.

[0050] In step S1 of this embodiment, the road management data includes four categories: road condition characteristics, traffic load parameters, climate data, and structural attributes, specifically:

[0051] Road surface condition characteristics: Road surface condition index (PCI) of the previous year;

[0052] Traffic load parameters: Annual average daily traffic volume (AADT), annual average daily freight traffic volume (AADTT).

[0053] Climate data: Total number of high-temperature days (≥35℃), consecutive high-temperature days (3 or more consecutive days with an average daily temperature > 35℃ in a year), total number of low-temperature days (≤0℃), consecutive low-temperature days (3 or more consecutive days with an average daily temperature < 0℃ in a year), annual rainfall;

[0054] Structural attributes: base layer thickness, base layer material type, surface layer thickness, surface layer material type, road age.

[0055] The multi-source feature preprocessing method is as follows:

[0056] Continuous variable standardization: Z-score standardization is used for continuous variables such as traffic volume (AADThe and AADTT), climate data (number of days with high temperature and low temperature, rainfall), pavement structure parameters (base layer and surface layer thickness), historical PCI data, and road age to eliminate dimensional differences.

[0057] One-hot encoding of categorical variables: Categorical variables such as base material type and surface material type are converted into binary vectors with one-hot encoding (e.g., cement-stabilized crushed stone is encoded as [1,0,0]) to avoid priority bias in categorical numericalization;

[0058] Road age is stored separately: Standardized parameters for road age are stored separately for subsequent destandardization in location coding. Standardization is necessary because road age is a continuous feature, and its numerical range may differ by magnitude from other features (such as rainfall, PCI values, etc.). Standardization eliminates this magnitude difference. Destandardization is used to recover the physical meaning for location coding.

[0059] In this embodiment, the physical information Transformer model structure is as follows: Figure 2 As shown, it specifically includes:

[0060] Input projection layer: Employs a fully connected network to map the road feature vectors obtained from multi-source feature preprocessing to the Transformer model dimension. The mapped road surface feature vector is obtained. ;

[0061] Physically Guided Road Age Location Coding Layer: Targeting the temporal decay pattern of road surface characterized by "slow early decay and rapid mid-term decay," a location coding system is designed for road age to explicitly capture the driving effect of road age on condition decay. Road age-aware location coding yields a road age location coding vector. This vector is then fused with the mapped road surface feature vector to obtain a road surface feature vector incorporating road age information, thereby capturing the age-driven exponential decay pattern of road surface condition.

[0062] First, reverse the standardization of road age, and then convert the standardized road age... The original road age can be restored using the following formula:

[0063] ;

[0064] Where t is the original road age. The standard deviation of road age, This represents the average road age.

[0065] Encoding Vector Generation: Based on the original road age t, a sine-cosine function is used to generate the road age location encoding vector. The formula is:

[0066] ;

[0067] ;

[0068] in, The road age location code value is calculated based on the sine function. The road age location code value is calculated based on the cosine function. For the physical information Transformer model dimension, For encoding dimension indexes.

[0069] Feature fusion: Encoding road age location vectors With road surface feature vector By concatenating the data, we obtain a pavement feature vector that incorporates road age information:

[0070] ;

[0071] in, To integrate road surface feature vectors with road age information, This is a road feature vector obtained by performing multi-source feature preprocessing on road management data.

[0072] The Transformer encoder consists of two stacked encoder layers. Each encoder layer contains a multi-head self-attention network and a feedforward network. After processing by the multi-head self-attention network, residual connections and normalization are performed. After processing by the feedforward network, residual connections and normalization are also performed before outputting to the next stage. The multi-head self-attention network has 4 attention heads and calculates the global dependency weights between features to capture the collaborative influence of multi-source features. The feedforward network adopts a "linear layer + ReLU activation + linear layer" structure.

[0073] Output layer: Maps the features extracted by the Transformer encoder to normalized road condition index (PCI) predictions. And then, through destandardization, it is restored to the original scale prediction value. Simultaneously, physical information constraints are incorporated to ensure the engineering rationality of the results. The inverse standardization formula is:

[0074] ;

[0075] in, The standard deviation of the road surface condition index. This represents the average value of the road surface condition index.

[0076] Physical Information Constraint Module: The physical information constraint module introduces physical laws into the Transformer model through road condition attenuation equations and interval constraints, avoiding predictions that do not conform to engineering common sense caused by purely data-driven approaches.

[0077] Through the composite loss function Collaborative optimization of learnable parameters related to physical constraints in Transformer model parameters ( , , This allows the Transformer model to both fit the data and adhere to physical laws. (Composite loss function) Loss due to data fitting Physical law constraint loss Deviation upper limit penalty loss The composition, and the formula are:

[0078] ;

[0079] in, The loss weights are constrained by physical laws; The upper limit of the deviation penalty loss weight is determined by the interval constraint;

[0080] Calculated based on the mean squared error between the predicted and actual values ​​from the Transformer model; Calculated based on the mean square error between the predicted and actual values ​​of the road surface condition degradation equation.

[0081] The predicted value of the road surface condition attenuation equation is calculated based on the following formula:

[0082] ;

[0083] in, This refers to the road surface condition index. The initial value of the road condition index is the road condition index value at t=0. In this embodiment, we take... =100; t is the road age; For learnable parameters related to the magnitude of initial decay of road surface conditions, For learnable parameters related to the rate of road surface condition decay, These are learnable parameters related to the stage characteristics of pavement decay, which affect the shape of the decay curve.

[0084] Deviation cap penalty loss The calculation formula is:

[0085] ;

[0086] Where N is the number of samples, For the first The original scale pavement condition index prediction value of each sample is obtained by inverse standardization of the pavement condition index output by the physical information Transformer model; For the first Each sample is a predicted value based on the pavement condition attenuation equation; upper limit penalty loss for deviation. It is a penalty function designed to prevent model predictions from deviating significantly from physical laws. The function ensures that a penalty is only applied when the deviation exceeds 4.5 (an engineering experience threshold, corresponding to a reasonable fluctuation range for the PCI score), and the penalty is 0 for deviations below the limit. This approach neither suppresses reasonable deviations nor fails to effectively constrain extreme deviations.

[0087] The collaborative optimization and prediction process is as follows:

[0088] The physical information Transformer model parameters are combined with the road condition attenuation equation parameters ( , , All parameters are set to learnable parameters. Optimizer selection: Adam optimizer is used; learning rate is set to 0.001; batch size is 32.

[0089] Iterative optimization: Input the training set samples into the Transformer model of physical information, calculate the gradient of the parameters through backpropagation, and update the parameters along the negative direction of the gradient. The formula is:

[0090] ;

[0091] in, For the learnable parameters in the k-th iteration, For the learnable parameters in the (k+1)th iteration, For learning rate, The gradient of the parameter.

[0092] After preprocessing the pavement management data of the road surface to be predicted in step S1, the data is input into the optimal physical information Transformer model, which outputs the predicted value of the Road Condition Index (PCI) and the parameters of the pavement condition attenuation equation. , , ), to complete the prediction of road surface condition degradation.

[0093] After optimization, The output value is 0.3, which reflects the proportional characteristics of the initial attenuation of the road surface condition—the engineering understanding that the early condition loss of a new road surface will not be too large; The output value is 0.7. This result accurately describes the nonlinear characteristics of the road surface condition degradation rate, indicating that the actual degradation process is not uniform, but rather shows a phased change that is fast at first and then slows down with the increase of road age. The output value is 4, which reflects the curve characteristics of the road surface condition transitioning from the rapid decay period to the stable decay period.

[0094] This paper utilizes the multi-head self-attention mechanism of the Transformer encoder to capture the complex global dependencies among multi-source features. Furthermore, a composite loss function incorporating physical constraints is designed. This function not only measures the error between the predicted and actual values ​​but also forces the prediction results to conform to the classical road condition degradation equation and penalizes deviations exceeding engineering experience thresholds, ensuring the physical rationality of the model. The parameters of the Transformer model based on physical information are collaboratively optimized using the backpropagation algorithm to minimize the composite loss function, ultimately achieving accurate road condition predictions that both highly fit the data and strictly adhere to physical laws.

[0095] To better demonstrate the advanced nature of this invention, the most mature baseline models in the field of pavement condition prediction, the Artificial Neural Network (ANN) model and the Transformer model, are used to construct pavement condition decay prediction models, respectively, in contrast to the Transformer model based on physical information in this invention. The evaluation metrics are the coefficient of determination (R²), mean squared error (MSE), and mean absolute error (MAE). A comparison of the ANN model's predicted values ​​with the actual values ​​is shown below. Figure 3 The accuracy was 0.908, the MSE was 0.089, and the MAE was 0.246; the comparison between the Transformer model's predicted values ​​and the actual values ​​was as follows: Figure 4 The accuracy was 0.927, the MSE was 0.073, and the MAE was 0.185; the comparison between the Transformer model's predicted and actual values ​​for physical information was as follows: Figure 5 The accuracy was 0.943, the MSE was 0.062, and the MAE was 0.183. Its coefficient of determination was, for example... Figure 6 Comparison of mean square error and mean absolute error Figure 7 Compared with the classic ANN model, the physical information Transformer model improves R² by 3.5 percentage points, reduces MSE by 30.3%, and reduces MAE by 25.6%. Compared with the baseline Transformer model, R² increases by another 1.6 percentage points, MSE decreases by another 15.1%, and MAE decreases by another 1.1%, comprehensively refreshing the upper limit of model performance and fully verifying the advanced nature of this invention.

[0096] To deeply analyze the decision-making logic of the Physical Information Transformer model in predicting the Road Condition Index (PCI) and provide interpretive support for road maintenance decisions, the SHAP interpretability framework is adopted. The specific process is as follows:

[0097] Reference Figure 1 SHAP value calculation: Input the test set samples into the trained Transformer model of physical information, calculate the SHAP value of each feature such as road age, previous year's PCI, AADTT, etc., for the prediction results of each sample, and quantify the marginal contribution of the features; Feature contribution visualization: Rank the importance of features, draw feature importance analysis charts, and show the distribution of SHAP values ​​of each feature; It can provide support for maintenance decision making: Based on the ranking of SHAP values, identify key influencing factors and provide a quantitative basis for the division of maintenance priorities.

[0098] First, an interpretability analysis is performed on the road condition decay prediction model based on the Transformer model, such as... Figure 8 Then, an interpretability analysis is performed on the road condition decay prediction model based on the Transformer model of physical information, such as... Figure 9Without physical information, the contribution percentage was 51.59%, which increased to 59.52% after adding physical information. This indicates that physical constraints reinforce the role of road age as the core driving factor for pavement condition degradation, which is more in line with actual engineering experience. The contribution percentage of the previous year's PCI without physical information was 14.41%, which decreased to 10.93% after adding physical information. This is because the model does not need to rely excessively on the pure data feature of the previous year's condition; the physical information Transformer model can more efficiently deduce the current condition through road age and physical degradation equations. Other auxiliary features (such as base course material, AADTT, temperature days, etc.): their contribution percentages generally decreased slightly. This indicates that physical information makes the model more dependent on road age and physical degradation patterns, compressing the relative importance of auxiliary features and allowing the model to focus more on core physical driving factors.

[0099] In summary, the injection of physical information effectively guides the model to focus on the age-dependent decay core mechanism, which not only strengthens the contribution of core physical factors, but also makes the distribution of feature influences more in line with engineering laws, ultimately improving the interpretability and physical rationality of the model.

[0100] Another embodiment of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the road condition degradation prediction method that integrates road age perception and physical constraints.

[0101] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the road condition decay prediction method that integrates road age perception and physical constraints.

[0102] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for predicting pavement condition degradation by integrating road age perception and physical constraints, characterized in that, Includes the following steps: S1. Collect pavement management data of unmaintained road sections and perform multi-source feature preprocessing on the pavement management data to obtain pavement feature vectors. All pavement management data are processed to obtain different pavement feature vectors to form a dataset. S2. Based on the Transformer model, a physical information constraint module is introduced. The physical information constraint module introduces physical laws into the Transformer model through the road condition decay equation and interval constraints, thereby constructing a physical information Transformer model. The dataset is divided into a training set and a test set. The physical information Transformer model is trained using the training set to obtain the optimal hyperparameters. The test set is used for testing and verification. If the verification is successful, the optimal physical information Transformer model for road condition decay prediction is obtained. S3. Introduce the SHAP interpretability framework to perform interpretability analysis on the optimal physical information Transformer model; The physical information Transformer model includes: Input projection layer: A fully connected network is used to map the road feature vector obtained from multi-source feature preprocessing to the Transformer model dimension, resulting in the mapped road feature vector; Physically guided road age location coding layer: Road age sensing location coding is performed to obtain road age location coding vector. The road age location coding vector is then fused with the mapped road surface feature vector to obtain a road surface feature vector with fused road age information. Transformer encoder: It consists of two stacked encoder layers. Each encoder layer contains a multi-head self-attention network and a feedforward network. After processing by the multi-head self-attention network, residual connections and normalization are performed. After processing by the feedforward network, residual connections and normalization are also performed before outputting to the next stage. Output layer: The features extracted by the Transformer encoder are mapped to the normalized road condition index (PCI) prediction value, and then denormalized to restore the original scale prediction value, while introducing physical information constraints; Physical Information Constraint Module: The physical information constraint module introduces physical laws into the Transformer model through road condition attenuation equations and interval constraints; Through the composite loss function The learnable parameters of the Transformer model, which are related to physical constraints, are collaboratively optimized to ensure that the Transformer model both fits the data and follows physical laws. This is achieved through a composite loss function. Loss due to data fitting Physical law constraint loss Deviation upper limit penalty loss The composition, and the formula are: ； in, The loss weights are constrained by physical laws; The upper limit of the deviation penalty loss weight is determined by the interval constraint; Calculated based on the mean squared error between the predicted and actual values ​​from the Transformer model; Calculated based on the mean square error between the predicted and actual values ​​of the road surface condition attenuation equation.

2. The road surface condition degradation prediction method according to claim 1, characterized in that, The physical guidance road age location coding layer processing includes: First, road age is destandardized, and the original road age is restored based on the standardized road age. Encoding vector generation: Based on the original road age, a sine-cosine function is used to generate the road age location encoding vector.

3. The road surface condition degradation prediction method according to claim 1, characterized in that, The predicted value of the road surface condition attenuation equation is calculated based on the following formula: ； in, This refers to the road surface condition index. This is the initial value of the pavement condition index, i.e., the pavement condition index value at t=0, where t is the pavement age; For learnable parameters related to the magnitude of initial decay of road surface conditions, For learnable parameters related to the rate of road surface condition decay, These are learnable parameters related to the stage characteristics of pavement condition decay.

4. The road surface condition degradation prediction method according to claim 1, characterized in that, The upper limit of the deviation penalty loss The calculation formula is: ； Where N is the number of samples, For the first The original scale pavement condition index prediction value of each sample is obtained by inverse standardization of the pavement condition index standardized value output by the physical information Transformer model; For the first Each sample is a predicted value based on the road condition attenuation equation.

5. The method for predicting road surface condition degradation according to claim 1, characterized in that, The multi-source features include four categories: road surface condition features, traffic load parameters, environmental impact factors, and structural attributes, specifically: Road surface condition characteristics: Road surface condition index value of the previous year; Traffic load parameters: Annual average daily truck traffic volume, annual average daily traffic volume; Environmental impact factors: Total number of high temperature days, consecutive high temperature days, total number of low temperature days, consecutive low temperature days, rainfall; Structural attributes: Base course thickness, base course material type, surface course thickness, surface course material type, road age.

6. The method for predicting road surface condition degradation according to claim 1, characterized in that, The multi-source feature preprocessing method is as follows: standardize the continuous variables in the multi-source features; perform one-hot encoding on the categorical variables in the multi-source features; and perform separate standardization on the road age features.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images