Method for predicting performance of asphalt pavement based on attention mechanism fusion of LSTM and ELM

Abstract

The present application relates to the field of highway maintenance and pavement performance prediction technology, and specifically discloses a method for predicting the performance of asphalt pavement based on the fusion of LSTM and ELM with an attention mechanism, which comprises the following steps: S1, obtaining multi-source original detection data, and obtaining a time series feature set and a static feature set through preprocessing and feature extraction; S2, constructing an attention-LSTM-ELM fusion prediction model, inputting the time series feature set into the LSTM branch to extract time series dependent features of pavement performance degradation, inputting the static feature set into the ELM branch to extract static inherent features of pavement, and adaptively distributing weights by the attention fusion layer to weight and fuse the two output feature vectors; S3, training the prediction model using a Huber loss function; and S4, synchronously predicting four pavement performance indicators, namely, RD, SFC, TCCI and IRI, using the trained prediction model. The method can simultaneously mine the time series evolution and static inherent features of pavement, has good robustness to outliers, can realize multi-index synchronous prediction, and provides a basis for highway asphalt pavement maintenance decision-making.

Description

Technical Field

[0001] This invention relates to the field of highway maintenance and pavement performance prediction technology, and in particular to a method for predicting the service performance of asphalt pavement based on the fusion of LSTM and ELM using an attention mechanism. Background Technology

[0002] In the context of large-scale highway maintenance, asphalt pavement performance prediction has become a crucial support for road maintenance decision-making. The service condition of highway asphalt pavements directly affects driving safety, ride comfort, and the overall operational efficiency of the road network. Its performance degradation is influenced by multiple factors, including pavement structural parameters, traffic loads, and climate conditions, exhibiting significant strong nonlinearity and dynamic temporal characteristics. Therefore, accurately predicting pavement performance is key to developing scientific maintenance plans, extending pavement service life, and reducing total life-cycle operating costs.

[0003] Currently, asphalt pavement performance prediction methods mainly fall into three categories: First, models based on mechanics-empirical theory, with the AASHTO model as a typical example. These models rely on complex mechanical theory derivations to construct a prediction system, exhibiting extremely high sensitivity to pavement material parameters and service boundary conditions. However, in practical engineering, the difficulty in parameter acquisition and low accuracy make it difficult to guarantee the prediction accuracy of such models in real-world applications. Second, traditional statistical models (such as regression analysis / grey theory). These models can only fit simple performance degradation trends and are unable to characterize the nonlinear coupling relationships and time-series dependencies among multiple factors. Third, single machine learning models. Most current machine learning models are built on limited-scale pavement performance datasets, resulting in limited generalization ability and difficulty in adapting to large-scale and differentiated measured data scenarios in road networks. When dealing with large-scale real-world data, they suffer from insufficient prediction stability and accuracy. Furthermore, most current prediction models fail to effectively integrate multi-source heterogeneous feature data such as traffic load, meteorological environment, pavement structure, and historical performance testing, resulting in single feature dimensions and insufficient information mining. In summary, existing machine learning methods still have significant shortcomings in accurately modeling complex temporal features and deeply mining multi-source data correlations, making it difficult to meet the engineering requirements for high-precision road performance prediction. Summary of the Invention

[0004] This invention provides a method for predicting the service performance of asphalt pavements based on an attention mechanism that integrates LSTM and ELM, addressing the problems existing in the prior art. This invention extracts the time-dependent features of pavement performance degradation using LSTM, optimizes the nonlinear mapping of high-dimensional features using ELM, and deeply couples the two types of features to achieve high-precision prediction of four indicators: RD, SFC, TCEI, and IRI, providing reliable technical support for pavement maintenance decisions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for predicting the service performance of asphalt pavements based on an attention mechanism that integrates LSTM and ELM, comprising the following steps: S1. Obtain multi-source raw detection data of highway asphalt pavement, including pavement structure data, environmental data, traffic data, and pavement performance index data; preprocess the multi-source raw detection data to obtain a regular sample dataset; extract features from the sample dataset to obtain a time-series feature set and a static feature set, wherein the time-series feature set is derived from historical environmental data and traffic data, and the static feature set is derived from pavement structure information and pavement performance index data for the predicted baseline year; S2. Construct an attention-LSTM-ELM fusion prediction model. This model has a dual-input structure, including an LSTM branch, an ELM branch, and an attention fusion layer. The LSTM branch input is a time series feature set. The LSTM forget gate, input gate, and output gate are used to extract the time series dependence features brought about by traffic accumulation effect and environmental change, and output a high-dimensional time series feature vector and the corresponding time series prediction results. The ELM branch input is a static feature set, which is used to mine the nonlinear correlation between pavement materials, structure and current performance, and outputs a high-dimensional static feature vector and the corresponding static prediction results. The attention fusion layer adaptively generates weight coefficients by combining the feature vectors output from the two branches, and adaptively weights and fuses the temporal prediction results and static prediction results to obtain a comprehensive feature vector, which is then output as a performance prediction result via the output mapping layer. S3. Model Training and Validation: The sample dataset is divided into training samples and validation samples according to road segment units. Dual-branch joint supervised training is conducted using the static feature set, temporal feature set, and ground truth values ​​of the road surface performance corresponding to the training samples. During training, the weight and bias parameters from the input layer to the hidden layer of the ELM branch are fixed and not iteratively updated after initialization. The output layer weights of the ELM branch are obtained through generalized inverse analytical solution. Only the parameters of the LSTM branch, attention fusion layer, and fused output layer are iteratively updated in reverse. The model prediction error is monitored using validation samples, and an early stopping strategy is used to suppress overfitting. The trained prediction model is obtained through iterative convergence. S4. The trained attention-LSTM-ELM fusion prediction model is used to predict the performance of the road surface, and four performance indicators are predicted simultaneously: road rutting RD, skid resistance SFC, transverse crack TCEI and smoothness IRI.

[0006] Further, in step S1, the pavement structure data includes the type and thickness parameters of the surface layer, base layer, and subbase layer; the environmental data includes the percentage of days with low temperatures, the percentage of days with high temperatures, and the percentage of days with rainfall; the traffic data includes the annual average daily traffic volume (AADT), passenger-to-freight ratio, and equivalent single-axle load (ESAL); and the pavement performance index data includes measured data of rutting (RD), skid resistance coefficient (SFC), transverse crack evaluation index (TCEI), and international roughness index (IRI).

[0007] Furthermore, in S1, the preprocessing includes numerical conversion of categorical features, missing value imputation, outlier removal and correction, and feature standardization; digital conversion of non-numerical categorical features is performed to unify the data format and dimensions of multi-source heterogeneous original data, and to obtain a sample dataset with consistent dimensions and compliant quality.

[0008] Furthermore, in order to achieve simultaneous prediction of the four indicators, 1000m is used as a unified evaluation unit. The unit mean and maximum value of the RD, IRI and SFC indicators collected at 100m intervals are statistically summarized. The TCEI indicator adopts the original 1000m evaluation scale.

[0009] Furthermore, in S1, the temporal feature set is a 10-dimensional temporal feature set, which is reconstructed from environmental data and traffic data from several consecutive years in chronological order to form a three-dimensional temporal tensor adapted to the input of the LSTM branch; the static feature set is an 8-dimensional static feature set, which is taken from pavement structure information and pavement performance index data of the prediction reference year.

[0010] Furthermore, in S2, the LSTM branch adopts a two-layer stacked network structure and has a built-in Dropout layer to suppress overfitting; the LSTM branch performs time-dependent feature extraction on the input time-series feature set, outputs a high-dimensional time-series feature vector, and outputs the road performance time-series prediction result through a fully connected regression layer for dimensional mapping.

[0011] Furthermore, in S2, the ELM branch adopts a single-layer feedforward network structure, fixes the input hidden layer parameters to stabilize the static feature extraction process; the ELM branch performs nonlinear correlation mining on the input static feature set, outputs a high-dimensional static feature vector, and solves the output layer weights through generalized inverse analysis to output the static prediction result of road performance.

[0012] Furthermore, in S2, the input of the attention fusion layer is the high-dimensional static feature vector output by the ELM branch and the high-dimensional temporal feature vector output by the LSTM branch. The fusion weight is adaptively generated through feature concatenation and weight calculation, and the proportion of the prediction results of the two branches is dynamically adjusted to output the final road performance fusion prediction result.

[0013] Furthermore, in S3, during model training, the Huber loss function is used to calculate the prediction error, the AdamW optimizer with weight decay is used to update the network parameters, and the training rate is controlled by exponentially decaying learning rate; hyperparameters are optimized through grid search, and training is stopped early based on the change in MSE loss on the validation set.

[0014] Furthermore, in S3, a 10-fold cross-validation method is used for model training and performance verification: all road segment samples are divided into 10 non-overlapping sample subsets according to road segment number, and 9 sets are selected in turn as training samples and the remaining set is selected as validation samples to complete 10 rounds of training and validation cycles; throughout the process, samples of the same road segment are constrained to belong to only a single subset to avoid data leakage; the average value of the 10 rounds of validation results is used as the model's overall prediction accuracy, and the root mean square error and coefficient of determination are used as the core evaluation indicators.

[0015] Furthermore, in S4, based on the predicted values ​​of RD, SFC, TCEI, and IRI output by the model, and in conjunction with the "Highway Technical Condition Assessment Standard" JTG5210-2018 and the TCEI special evaluation standard, the performance of asphalt pavement on expressways is divided into five technical evaluation levels: excellent, good, medium, inferior, and poor.

[0016] The beneficial effects achieved by one or more of the above technical solutions are as follows: (1) This application effectively captures the long-term temporal dependence features of road performance degradation by using LSTM, efficiently realizes the nonlinear mapping of high-dimensional features by using ELM, and combines the advantages of LSTM and ELM to solve the problem that a single model is difficult to take into account both temporal and nonlinear characteristics, thereby improving the prediction accuracy.

[0017] (2) Existing technologies mostly focus on a single performance index, and the prediction of a single index is difficult to support the formulation of scientific maintenance plans. This invention addresses the actual needs of engineering and achieves simultaneous prediction of four indicators: RD, SFC, TCEI, and IRI. Through large-scale experimental sample verification, it has achieved excellent prediction accuracy. The accuracy of each indicator is significantly better than traditional models such as BP neural network and XGBoost, as well as single models such as pure LSTM and pure ELM.

[0018] (3) This invention integrates data from four dimensions: pavement structure, environmental conditions, traffic load, and historical performance, and constructs a comprehensive analysis framework covering "structure-environment-load-state". This system overcomes the prediction bias caused by the one-sided input information of traditional models, and makes the prediction results more consistent with the actual degradation mechanism of pavement performance. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of the asphalt pavement performance prediction method of the present invention; Figure 2 This is a schematic diagram of the attention-LSTM-ELM fusion prediction model of the present invention; Figure 3 This is a detailed flowchart of the asphalt pavement service performance prediction method of the present invention; Figure 4 This is a scatter plot comparing the true value-predicted value fitting of the model of this invention with each comparative model; Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Figures 1-3 As shown, the asphalt pavement service performance prediction method based on the attention mechanism integrating LSTM (Long Short-Term Memory Network) and ELM (Extreme Learning Machine) includes the following steps: S1. Obtain multi-source raw detection data of highway asphalt pavement, including pavement structure data, environmental data, traffic data, and pavement performance index data; preprocess the multi-source raw detection data to obtain a regular sample dataset; extract features from the sample dataset to obtain a temporal feature set and a static feature set, respectively. The temporal feature set is derived from historical environmental and traffic data, and the static feature set is derived from pavement structure information and pavement performance index data for the prediction reference year; the temporal feature set is a 10-dimensional temporal feature set, which is reconstructed from historical environmental and traffic data over several consecutive years in chronological order to form a three-dimensional temporal tensor adapted to the input of the LSTM branch; the static feature set is an 8-dimensional static feature set, which is taken from pavement structure information and pavement performance index data for the prediction reference year.

[0024] The specific methods for acquiring and preprocessing the data are as follows: S11. Data Preparation: A database was built using historical monitoring data of 29 expressways in multiple provinces in East China owned by the project team; the long-term monitoring database of expressways independently accumulated by the project team was used as the data source, and four types of measured performance indicators were selected as the basic data for modeling. S12. Data Acquisition: To achieve simultaneous prediction of multi-dimensional pavement performance, pavement structure information (surface layer, base layer, subbase layer type and thickness) is systematically extracted from the project team's own database; environmental data (percentage of days with minimum temperature below 0℃, percentage of days with maximum temperature above 35℃, percentage of rainy days); traffic data (annual average daily traffic volume (AADT), passenger-freight ratio, equivalent single-axle load (ESAL)); and performance index data (IRI, RD, SFC, TCEI). S13. Perform data numerical conversion, data quality verification, and data cleaning operations on the acquired multi-dimensional data in sequence, followed by feature selection; specifically including: First, the pavement structure of each road segment and the measured values ​​of RD, SFC, TCEI, and IRI are extracted. Then, the monitoring data of asphalt pavement of highways of different traffic levels are extracted, and temperature and precipitation-related features are used as time series features. To achieve simultaneous prediction of the four indicators, a unified 1000m basic evaluation unit is adopted. For the RD, IRI, and SFC indicators with an original detection interval of 100m, statistical summaries are performed within each 1000m unit (such as taking the mean, maximum, or representative value); TCEI, on the other hand, retains the original 1000m evaluation scale.

[0025] To address the characteristics of different types of variables, a hierarchical encoding strategy can be adopted. For nominal variables with a small state cardinality, such as up / down traffic and lane type, one-hot encoding can be used to convert them into numerical features. For example, "up / down traffic + one lane" can be encoded as [1,0,1,0,0], thus avoiding the misleading effect of ordinal relationships between categories. For textual features such as road surface structure (AC, AK, SMA, etc.), a standardized vocabulary can be constructed first through word segmentation, combination, and deduplication, and then the text sequence can be mapped into a numerical vector. To address the issue of inconsistent vector lengths caused by differences in the number of defects in different road sections, a padding operation with trailing zeros can be used to align feature dimensions. For example, the mapped vector is [1,19,3,17,13,23,0,0,0], ensuring the consistency of data input.

[0026] For missing values, considering both data type and missing features, linear interpolation / spline interpolation / interpolation between adjacent years is preferred for continuous time series variables such as road surface performance, traffic volume, and climate. For non-time series numerical variables, KNN or median is used for imputation, while for categorical variables, mode is used for imputation or an "unknown" category is added. Outliers are identified using the IQR method and verified in conjunction with engineering experience. Obvious errors or measurement anomalies are removed, while extreme samples with practical engineering significance are retained or truncated to effectively reduce the interference of outliers on model training and ensure data quality and model prediction accuracy. For numerical features such as IRI, RD, and SFC, Min-Max normalization is used to scale them to the [0,1] interval. This normalization process eliminates the differences in scale and numerical magnitude between different features, preventing high-magnitude features from overly dominating the model learning process. For continuous variables with large scale differences, Z-Score standardization is used to eliminate the influence of different scales on model training.

[0027] After data preprocessing, the preprocessed data is used to construct an attention-LSTM-ELM fusion model for implementing the prediction method.

[0028] S2. Construct an attention-LSTM-ELM fusion prediction model. This model has a dual-input structure, including an LSTM branch, an ELM branch, and an attention fusion layer. The prediction model mainly uses ELM to receive basic information features of highway asphalt pavement and basic features of pavement performance as input, and learns material properties and implicit influencing factors of current performance. LSTM receives traffic load time series features and climate condition time series features as input, and captures the dynamic dependence of traffic load cumulative effect and climate change through forget gate, input gate, and output gate. The weights of ELM output and LSTM output are dynamically allocated through the sigmoid activation function to achieve deep coupling of the two types of features, and finally output the predicted values ​​of each performance.

[0029] The LSTM branch adopts a two-layer stacked network structure and has a built-in Dropout layer to suppress overfitting. It uses the forget gate, input gate, and output gate of LSTM to extract the time-dependent features brought about by traffic accumulation effect and environmental change. The input of the LSTM branch is a time-series feature set. The LSTM branch extracts time-dependent features from the input time-series feature set, outputs a high-dimensional time-series feature vector, and outputs the time-series prediction results of road performance through a fully connected regression layer for dimensional mapping.

[0030] The specific processing procedure for inputting temporal features into the LSTM branch is as follows: The LSTM branch input is a three-dimensional tensor with dimensions (Batch_Size, T, X); Batch_Size is the batch size during training, set to 32, and T is the time step, i.e. the number of historical years used for prediction. Grid search verification shows that the model performs best when T=8. Using the structural, traffic, climate, and performance detection data available at the end of year t as input, the pavement performance indicators for year t+1 are predicted.

[0031] A dual-layer stacked LSTM structure extraction unit extracts key dynamic information from time-series data. The input (10 dimensions) consists of time-series features, including AADT, ESAL, the proportion of days with minimum temperatures below 0°C, the proportion of days with maximum temperatures above 35°C, and the proportion of rainy days. This unit aims to capture the long-term dependence of traffic load accumulation on climate cycle changes.

[0032] The first LSTM layer has 128 units, and the input tensor dimension is (32, 8, 10). The computation of each LSTM unit at each time step involves a forget gate f. t Input gate i t Candidate cell status Cellular state C t and output gate o t This layer outputs all hidden states at 8 time steps, with dimensions (32, 8, 128), and the mathematical expressions are as follows: ; ; ; ; ; In the formula, f t i t o t These represent the outputs of the forget gate, input gate, and output gate, respectively, with dimensions (32, 128), consistent with the number of LSTM units; It represents the candidate cell state, with dimensions (32, 128). and These correspond to the cell states at the current time step and the previous time step, respectively, and serve as the core memory carrier of LSTM, with dimensions of (32, 128). and These are the hidden states of the current time step and the previous time step (i.e., the final output of the LSTM unit), with dimensions (32, 128). The input at time step t has dimensions (32, 10), which corresponds to the feature dimension of the input tensor of 10. These are the weight matrices for the forget gate, input gate, candidate cell state, and output gate, respectively, with dimensions (128, 128+10), where 128 is the number of LSTM units and 10 is the number of input features; These are the bias vectors corresponding to each component, each with a dimension of 128; This represents the Sigmoid activation function, whose output range is between 0 and 1, and is used to control the information throughput. It is the Tanh activation function, with an output range of -1 to 1, used to normalize the state of candidate cells; This represents the concatenation operation between the hidden state of the previous time step and the input of the current time step, with the concatenated dimension being (32, 128+10). This represents element-wise multiplication (i.e., the Hadamard product), used to perform element-wise multiplication between vectors.

[0033] A Dropout layer with a dropout rate of 0.2 was set between the first and second LSTM layers. During training, 20% of the neuron connections were randomly masked to alleviate the model overfitting problem.

[0034] The second LSTM layer has 64 units, and its input is the output dimension of the first layer after Dropout processing (32, 8, 128). Only the hidden state h of the last time step (year t) is extracted. t The final output dimension of the time-series unit is (32, 64).

[0035] The ELM branch adopts a single-layer feedforward network structure, with fixed input hidden layer parameters to stabilize the static feature extraction process; the input of the ELM branch is a static feature set, which is used to mine the nonlinear correlation between road materials, structure and current performance, and outputs a high-dimensional static feature vector and the corresponding static prediction results.

[0036] Specific process of static feature input ELM branch: The ELM branch is a single-hidden-layer feedforward neural network structure. Its core advantage lies in its excellent training efficiency and ability to converge to the global optimum. The 8-dimensional input features of the ELM are static and current-state features, including pavement structure coding, road age, lane location, and measured values ​​of RD, SFC, TCEI, and IRI for year t (the most recent year before prediction). It can capture the deep nonlinear relationship between material properties, structural characteristics, and current performance status.

[0037] Hidden Layer: The number of neurons in the hidden layer was determined to be 256 through grid search optimization. The activation function used is the Sigmoid function, whose expression is: For a batch containing 32 samples, the hidden layer output matrix H is calculated as follows: ; In the formula, The input data matrix has dimensions (32, 8). The weight matrix from the input layer to the hidden layer has dimensions (8, 256). This matrix follows a normal distribution N(0, 0.01) during model initialization and remains fixed during training. This is the hidden layer bias vector with dimensions (1, 256), which is also fixed after initialization.

[0038] Output Layer: The output layer weight β has a dimension of (256, 4), corresponding to the four prediction targets: RD, SFC, TCEI, and IRI. This weight is not optimized through gradient descent iterations, but is directly solved analytically using the Moore-Penrose generalized inverse matrix, as shown in the formula: ; In the formula, Let H be the generalized inverse of the hidden layer output matrix H, and Y be the true label matrix of the performance index corresponding to the sample batch, with dimensions (32, 4). The above analytical solution process is executed after the forward propagation of each training batch is completed.

[0039] The input to the attention fusion layer is the high-dimensional static feature vector output by the ELM branch and the high-dimensional temporal feature vector output by the LSTM branch. The fusion weight is adaptively generated through feature concatenation and weight calculation, and the proportion of the prediction results of the two branches is dynamically adjusted to output the final road performance fusion prediction result.

[0040] To achieve adaptive adjustment of the contribution weights of LSTM (Laser-Time Dynamic Features) and ELM (Electronic-State Features) to the final prediction result, this invention sets up an attention fusion layer, the specific implementation process of which is as follows: h output by LSTM t The feature vector z is obtained by concatenating the 64-dimensional (64-dimensional) feature vector with the output matrix H (256-dimensional) of the ELM hidden layer to obtain the fused feature vector z = [h]. t [H] (320 dimensions), then the attention weights are calculated through a fully connected layer: ; Where σ is the Sigmoid function, Wa is a 1×320 weight matrix, ba is the bias scalar, and the final predicted value of the model is: ; In the formula, W is the static prediction output of the ELM branch. o h t +b o This is the time series prediction output of the LSTM branch through the fully connected regression layer.

[0041] Through this mechanism, the model can adaptively determine the contribution ratio of historical time series patterns and current static features under different prediction scenarios.

[0042] S3. Model Training and Validation: The sample dataset is divided into training samples (Train set) and validation samples (Test set) according to road segment units. Dual-branch joint supervised training is conducted using the static feature set, temporal feature set, and ground truth values ​​of road performance corresponding to the training samples. During training, the weight and bias parameters from the input layer to the hidden layer of the ELM branch are fixed and not iteratively updated after initialization. The output layer weights of the ELM branch are obtained through generalized inverse analytical solution. Only the parameters of the LSTM branch, attention fusion layer, and fused output layer are iteratively updated in reverse. The model prediction error is monitored using validation samples, and an early stopping strategy is used to suppress overfitting. The trained prediction model is obtained through iterative convergence.

[0043] The training strategy of this application is as follows: an early stopping strategy is adopted (if the verification lumped mean squared error (MSE) does not decrease for 5 consecutive rounds, training is terminated to avoid overfitting), the AdamW optimizer with weight decay (weight decay coefficient 0.001) is selected and an exponential decay learning rate is configured (initial value 0.001, decaying to 1 / 10 of the original value every 10 rounds); the hyperparameters are determined by grid search to determine the optimal combination, the search range including the number of neurons in the ELM hidden layer (64-256), the dimension of the LSTM hidden layer (32-128) and the attention weight decay coefficient (0.001-0.01).

[0044] Model validation phase: Comparative experiments were conducted using real datasets from the database. Three traditional models, namely BP neural network, support vector regression (SVR), and XGBoost, as well as LSTM-ELM fusion models using sigmoid and softmax attention activation functions respectively, were selected as control groups. The models were comprehensively evaluated from the two dimensions of prediction accuracy and training efficiency.

[0045] like Figure 4 The figure shows a scatter plot comparing the true value-predicted value fitting of the model of this invention with various comparative models. The figure includes sample distribution and model evaluation metrics, where Train represents the training set and Test represents the test set. The comparison results show that: SVR and ELM models have poor overall fitting performance; BP and XGBoost models have limited generalization ability and cannot uncover the temporal evolution patterns of road surfaces; the single LSTM model only utilizes temporal information and does not adequately utilize static inherent features; the LSTM-ELM (Sigmoid) model, which directly concatenates data, has high accuracy on the training set, but carries the risk of overfitting.

[0046] This invention employs an LSTM-ELM fusion model with Softmax attention, which adaptively allocates the weights of two feature paths through the attention mechanism, resulting in a more balanced distribution of training and testing samples. This model suppresses overfitting while ensuring high prediction accuracy, and its overall prediction performance is superior to other comparative models.

[0047] The input layer weights and biases of the ELM branch are fixed after random initialization; at the beginning of each fold of training, the output layer weights β are solved once based on the output matrix of the ELM hidden layer of the training set samples; during training, only the parameters of the LSTM branch, the attention fusion layer and the fused output layer are updated.

[0048] The AdamW optimizer was selected. This optimizer improves the generalization performance of the model by decoupling the weight decay mechanism. The weight decay coefficient was set to 0.001.

[0049] The Huber loss function is adopted, which combines the advantages of MAE and MSE, exhibiting robustness to outliers and good stability during training. The parameters... Set to 1.0, the expression is: ; In the formula, y This is the true label of road surface performance indicators (sample measured true values); This is the model's predicted output value for this sample; The absolute error between the predicted value and the actual value; The Huber loss threshold parameter is used in this scheme. =1; The loss value calculated using the Huber loss function.

[0050] An exponentially decaying learning rate is configured, with an initial learning rate of 0.001. Every 10 training epochs, the learning rate decays to 1 / 10 of its original value to achieve rapid convergence in the early stages of training and fine-tuning of parameters later. The validation set loss is continuously monitored; if the validation set loss does not decrease for five consecutive epochs, training is terminated, and the model parameters corresponding to the minimum validation loss are loaded to suppress overfitting. For key hyperparameters such as the number of LSTM layers and units, Dropout rate, number of neurons in the ELM hidden layers, initial learning rate, and batch size, a systematic optimization is performed within a preset range using a grid search method to ensure optimal model performance.

[0051] In the model validation phase, comparative experiments were conducted using real datasets from the database. Three traditional models—BP neural network, support vector regression (SVR), and XGBoost—were selected as control groups, along with a pure ELM model and a pure LSTM model using sigmoid and softmax attention activation functions, respectively. The models were comprehensively evaluated from two dimensions: prediction accuracy and training efficiency.

[0052] In the validation experiment, the evaluation metrics used were root mean square error (RMSE) and coefficient of determination (R²), which are commonly used metrics in regression analysis. These metrics respectively measure the average error between the predicted and actual values ​​and the proportion of variance explained by the model. For the... k Performance metrics ( k =1, 2, 3, 4 correspond to RD, SFC, TCEI, and IRI respectively), and the calculation formula is as follows: ; ;

[0053] in, Representing the i The first sample k Performance metrics ( k The true value of (RD / SFC / TCEI / IRI), Representing the i The first sample k Predicted values ​​of each performance indicator Representing all samples k The average of the true values ​​of each performance indicator n Represents the number of samples.

[0054] Performance prediction: The final prediction value of the LSTM-ELM model The formula is obtained by weighted summation of the preliminary prediction results from the two units: ; In the formula, This is the prediction output (static prediction result) for the ELM branch. This is the prediction output (time series prediction result) of the LSTM branch through the fully connected regression layer; through this mechanism, the model can adaptively determine the contribution ratio of historical time series patterns and current static-state features under different prediction scenarios.

[0055] S4. The trained attention-LSTM-ELM fusion model is used to predict the performance of the road surface, and four performance indicators are predicted simultaneously: road rutting RD, skid resistance SFC, transverse crack TCEI, and smoothness IRI.

[0056] Specifically: Dataset Partitioning and Prediction: Using measured values ​​of RD, SFC, TCEI, and IRI as modeling targets, 73,258 valid samples (after data cleaning) from 29 test sections of highways in multiple provinces of East China were selected. A 10-fold cross-validation method was used to train and validate these 73,258 valid samples, with approximately 90% of the samples in each fold serving as the training set and approximately 10% as the validation set. The data was then input into the LSTM-ELM feature fusion model for prediction. During dataset partitioning, cross-validation was performed using road segment numbers as grouping units to ensure that data from the same road segment within the same prediction window did not appear simultaneously in the training and validation sets, thus avoiding information leakage.

[0057] Error Analysis: The trained prediction model was used to predict and analyze 7326 validation set samples. The results show that the absolute value of the RD prediction error fluctuates around 0.5 mm, with a maximum error of 1.18 mm; 91.2% of the samples have an SFC error < 3; 89.7% of the samples have a TCEI error < 2; and 90.5% of the samples have an IRI error < 0.2 m / km. Only a few extreme samples have slightly higher errors, and the overall prediction performance of the model is excellent.

[0058] Grade Evaluation: Based on the calculation results of the "Highway Technical Condition Evaluation Standard" (JTG 5210-2018) and the TCEI evaluation method proposed by Zhou Lan, the technical grade evaluation of each performance index is carried out. The TCEI calculation formula is as follows; ; ; ; ;

[0059] In the formula, TCEI is the transverse crack evaluation index, ranging from 0 to 100, with 100 representing the best transverse crack condition of the road surface; TCS is the transverse crack spacing (unit: m); TWR is the transverse crack penetration, dimensionless, and if TWR is greater than 1, it is taken as 1; TCL is the transverse crack evaluation length (unit: m); TCN is the total number of transverse cracks in the evaluated road segment; TCTL is the total length of transverse cracks (unit: m); L is the length of the evaluated road segment (unit: m); and B is the lane width (unit: m).

[0060]

[0061] In the table above, under Rutting Condition (RD), no ruts (Excellent) indicates a smooth ride, slight ruts (Good) indicate a fairly smooth ride, noticeable ruts (Medium) indicate an unstable ride, severe ruts (Poor) indicate a very unstable ride, and extremely deep ruts (Very Unstable) indicate a very unstable ride. Under Skid Resistance (SFC), excellent (Excellent) indicates high safety assurance, good (Good) indicates that safety requirements are met, average (Medium) indicates a reduced safety margin, insufficient (Poor) indicates a potential safety hazard, and severely insufficient (Very Unstable) indicates a high risk. Under Smoothness (IRI), very smooth (Excellent) indicates excellent comfort, smooth (Good) indicates good comfort, slightly bumpy (Medium) indicates average comfort, uneven (Poor) indicates poor comfort, and severely uneven (Very Unstable) indicates extremely poor comfort. In the Crack Condition Index (TCEI), Excellent indicates almost no damage, Good indicates only minor cracks, Medium indicates noticeable cracks, Poor indicates severe cracks, and Very Poor indicates severe and dense cracks.

[0062] The grading criteria in the table above are used to evaluate the levels of RD, IRI, SFC, and TCEI.

[0063] This embodiment is only used to illustrate the technical solution of the present invention and does not limit the scope of protection.

[0064] The above specific embodiments should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, any alternative improvements or modifications made to the embodiments of the present invention shall fall within the scope of protection of the present invention.

[0065] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A method for predicting the service performance of asphalt pavement based on an attention mechanism that integrates LSTM and ELM, characterized in that, Includes the following steps: S1. Obtain multi-source raw detection data of highway asphalt pavement, including pavement structure data, environmental data, traffic data, and pavement performance index data; preprocess the multi-source raw detection data to obtain a regular sample dataset; extract features from the sample dataset to obtain a time-series feature set and a static feature set, wherein the time-series feature set is derived from historical environmental data and traffic data, and the static feature set is derived from pavement structure information and pavement performance index data for the predicted baseline year; S2. Construct an attention-LSTM-ELM fusion prediction model. This model has a dual-input structure, including an LSTM branch, an ELM branch, and an attention fusion layer. The LSTM branch input is a time series feature set. The LSTM forget gate, input gate, and output gate are used to extract the time series dependence features brought about by traffic accumulation effect and environmental change, and output a high-dimensional time series feature vector and the corresponding time series prediction results. The ELM branch input is a static feature set, which is used to mine the nonlinear correlation between pavement materials, structure and current performance, and outputs a high-dimensional static feature vector and the corresponding static prediction results. The attention fusion layer adaptively generates weight coefficients by combining the feature vectors output from the two branches, and adaptively weights and fuses the temporal prediction results and static prediction results to obtain a comprehensive feature vector, which is then output as a performance prediction result via the output mapping layer. S3. Model training and validation: The sample dataset is divided into training samples and validation samples according to road segment units. The static feature set, time series feature set and pavement performance true value corresponding to the training samples are used to carry out dual-branch joint supervised training. During training, the weight and bias parameters from the input layer to the hidden layer of the ELM branch are fixed. These parameters are not updated iteratively after initialization. The output layer weights of the ELM branch are obtained by generalized inverse analytical solution. Only the parameters of the LSTM branch, the attention fusion layer, and the fused output layer are updated iteratively in reverse. The model prediction error is monitored using validation samples, and an early stopping strategy is used to suppress model overfitting. The trained prediction model is obtained by iterative convergence. S4. The trained attention-LSTM-ELM fusion model is used to predict the performance of the road surface, and four performance indicators are predicted simultaneously: road rutting RD, skid resistance SFC, transverse crack TCEI, and smoothness IRI.

2. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 1, characterized in that, In S1, the pavement structure data includes the type and thickness parameters of the surface layer, base layer, and subbase layer; the environmental data includes the percentage of days with low temperatures, the percentage of days with high temperatures, and the percentage of days with rainfall; the traffic data includes the annual average daily traffic volume (AADT), passenger-to-freight ratio, and equivalent single-axle load (ESAL); and the pavement performance index data includes measured data of rutting (RD), skid resistance coefficient (SFC), transverse crack evaluation index (TCEI), and international roughness index (IRI).

3. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM as described in claim 1 or 2, characterized in that, In S1, the preprocessing includes categorical feature numerical transformation, missing value imputation, outlier removal and correction, and feature standardization. Non-numerical categorical features are digitized to unify the data format and dimensions of heterogeneous original data from multiple sources, resulting in a standardized sample dataset with consistent dimensions and compliant quality.

4. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 3, characterized in that, To achieve simultaneous prediction of the four indicators, 1000m was used as a unified evaluation unit. The mean and maximum values ​​of the RD, IRI, and SFC indicators collected at 100m intervals were statistically summarized. The TCEI indicator continued to use the original 1000m evaluation scale.

5. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 4, characterized in that, In S1, the temporal feature set is a 10-dimensional temporal feature set, which is reconstructed from environmental data and traffic data from several consecutive years in chronological order to form a three-dimensional temporal tensor adapted to the input of the LSTM branch; the static feature set is an 8-dimensional static feature set, which is taken from pavement structure information and pavement performance index data of the prediction reference year.

6. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 5, characterized in that, In S2, the LSTM branch adopts a two-layer stacked network structure and has a built-in Dropout layer to suppress overfitting; the LSTM branch extracts time-dependent features from the input time-series feature set, outputs a high-dimensional time-series feature vector, and outputs the time-series prediction results of road performance through a fully connected regression layer for dimensional mapping.

7. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 6, characterized in that, In S2, the ELM branch adopts a single-layer feedforward network structure, fixes the input hidden layer parameters to stabilize the static feature extraction process; the ELM branch performs nonlinear correlation mining on the input static feature set, outputs a high-dimensional static feature vector, and solves the output layer weights through generalized inverse analysis to output the static prediction result of road performance.

8. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 7, characterized in that, In S2, the input to the attention fusion layer is the high-dimensional static feature vector output by the ELM branch and the high-dimensional temporal feature vector output by the LSTM branch. The fusion weight is adaptively generated through feature concatenation and weight calculation, and the proportion of the prediction results of the two branches is dynamically adjusted to output the final road performance fusion prediction result.

9. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 8, characterized in that, In S3, during model training, the Huber loss function is used to calculate the prediction error, the AdamW optimizer with weight decay is used to update the network parameters, and the training rate is controlled by exponential decay learning rate; hyperparameters are optimized through grid search, and training is stopped early based on the change in MSE loss on the validation set.

10. The asphalt pavement service performance prediction method based on attention mechanism fusion of LSTM and ELM according to claim 9, characterized in that, In S3, a 10-fold cross-validation method is used for model training and performance verification: all road segment samples are divided into 10 non-overlapping sample subsets according to the road segment number. Nine sets are selected in turn as training samples and the remaining set is used as validation samples to complete 10 rounds of training and validation cycles. Throughout the process, samples of the same road segment are constrained to belong to only a single subset to avoid data leakage. The average value of the 10 rounds of validation results is used as the overall prediction accuracy of the model, and the root mean square error and coefficient of determination are used as the core evaluation indicators.

Images