A method for pavement structure modulus inversion and damage evaluation based on deflection slope

Abstract

This invention discloses a method for pavement structure modulus inversion and damage assessment based on deflection slope. The method first acquires measured deflection slope data for multiple pavement channels under traffic load and preprocesses the data. Then, a neural network forward model is constructed and trained based on pavement structure parameter combinations and corresponding theoretical deflection slopes to predict the deflection slope. Next, a fitness function is constructed with the objective of minimizing the relative root mean square error between the measured deflection slope and the predicted values ​​from the neural network forward model, and a genetic algorithm is used to find the optimal inversion solution for the modulus of each pavement structural layer within a set parameter space. Finally, based on the measured deflection slope data of the hard shoulder in the same road segment, the modulus of each structural layer of the hard shoulder is inverted and used as a health benchmark to calculate the damage index of the driving lane structural layer. This method effectively solves the problems of low forward calculation efficiency, susceptibility to local optima, and lack of damage quantification benchmarks in traditional methods.

Description

Technical Field

[0001] This invention relates to a method for inverting the pavement structure modulus and evaluating damage based on deflection slope, belonging to the field of non-destructive testing and structural health diagnosis technology in road engineering. Background Technology

[0002] The bearing capacity of pavement structure is the core of road service performance evaluation, and the modulus of each structural layer is a key parameter reflecting its mechanical state. Traditional modulus inversion methods mainly rely on falling weight deflectometers (FWD) for static or quasi-static testing, which has shortcomings such as low detection efficiency, need to interrupt traffic, and difficulty in covering road network scale. In recent years, high-speed laser deflection detection technology (such as TSD) can continuously acquire dynamic deflection slope data of the pavement at normal driving speeds, greatly improving detection efficiency. However, modulus inversion for such dynamic deflection slope data still faces three major technical bottlenecks:

[0003] Bottleneck in forward modeling efficiency: Traditional inversion methods require repeated calls to layered elasticity theory models or finite element models for forward modeling calculations, with each calculation taking seconds or even minutes, which cannot meet the needs of rapid or real-time inversion of massive amounts of detection data.

[0004] The problem of parameter ambiguity: Deflection basin curves exhibit a significant "different parameters, same effect" phenomenon, meaning that multiple combinations of structural layer moduli may produce extremely similar deflection responses. Traditional gradient-based optimization methods are prone to getting trapped in local optima and struggle to characterize and output reasonable ranges for the moduli, resulting in a lack of robustness and interpretability in the inversion results.

[0005] Damage assessment lacks a benchmark: existing methods mostly only output the absolute value of the modulus, which makes it difficult to objectively quantify the degree of fatigue damage to the pavement structure layer under repeated traffic loads, and lack intuitive damage assessment indicators for maintenance decision-making.

[0006] Therefore, there is an urgent need to develop a method for pavement structure modulus inversion and damage evaluation that is computationally efficient, has strong global optimization capabilities, can handle multiple solutions, and provides a quantitative damage evaluation benchmark. Summary of the Invention

[0007] The purpose of this invention is to provide a method for inverting the pavement structure modulus and evaluating damage based on deflection slope, so as to solve the problems of low forward calculation efficiency, easy getting trapped in local optima and lack of damage quantification benchmark in traditional methods.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for pavement structure modulus inversion and damage evaluation based on deflection slope includes the following steps: Step 1: Obtain the measured deflection slope data of multiple lanes of the road surface under traffic load, and preprocess the acquired data; Step 2: Use Latin hypercube sampling to generate multiple sets of pavement structure parameter combinations and calculate the theoretical deflection slope corresponding to each combination. Based on the pavement structure parameter combinations and the corresponding theoretical deflection slopes, construct and train a neural network forward surrogate model to predict the deflection slope. Step 3: With the goal of minimizing the relative root mean square error between the measured deflection slope and the predicted value of the neural network forward surrogate model, construct a fitness function and use a genetic algorithm to find the optimal solution for the inversion of the modulus of each structural layer of the pavement within the set parameter space. Step 4: Based on the measured deflection slope data of the hard shoulder in the same road segment, invert the modulus of each structural layer of the hard shoulder and use it as a health benchmark to calculate the damage index of the driving lane structural layer. .

[0010] Preferably, in step 1, the measured deflection slope data includes multiple deflection slope values ​​at different horizontal distances from the load center.

[0011] Preferably, data preprocessing includes outlier removal, jump detection, and missing value imputation; wherein, Outlier removal involves removing all data at the corresponding point when the deflection slope of any channel is less than 0 or greater than 200. Jump detection is used when the difference between adjacent detection points in the same channel exceeds 3 times the historical standard deviation, and the result is considered an anomaly and removed. Missing value imputation is used to fill in individual missing values ​​caused by device packet loss using linear interpolation.

[0012] Preferably, in step 2, the specific configuration for constructing and training the neural network forward surrogate model is as follows: A multilayer perceptron network with 5 hidden layers was constructed, with each hidden layer followed by a Swish activation function. The number of neurons in each layer was configured as 128, 128, 256, 256, and 128, respectively, and no activation function was set for the output layer.

[0013] The optimizer uses Adam and introduces an early stopping mechanism to prevent model overfitting.

[0014] Preferably, in step 2, the input feature preprocessing method of the multilayer perceptron network is as follows: The network input vector contains modulus parameters and thickness parameters, and the natural logarithm of the modulus parameters of all inputs is taken. All processed features are Z-score standardized.

[0015] Preferably, in step 2, the weighted root mean square error loss function used to train the neural network forward surrogate model is defined as:

[0016] In the formula, N is the number of samples; This refers to the number of sensor channels; Let be the weight of the i-th channel; Let be the deflection slope of the k-th sample and the i-th channel predicted by the neural network; Calculate the deflection slope for the k-th sample and the i-th channel according to the theory. Let the root mean square value of the theoretical deflection slope of all channels of the k-th sample satisfy:

[0017] Preferably, in step 3, the objective function of the genetic algorithm for modulus optimization is defined as:

[0018] In the formula, For the first Measured deflection slope of the channel The predicted value is from the neural network surrogate model. This represents the number of sensor channels.

[0019] Preferably, in step 3, the specific operating parameters of the genetic algorithm are configured as follows: Real number encoding was used, the population size was set to 1000, and population initialization was completed based on Latin hypercube sampling; The fitness function is defined as follows: ; Select the tournament selection with a scale of 3; The crossover operation is a simulated binary crossover, with crossover probability... Distribution index ; The mutation operation is a polynomial mutation, with a mutation probability of... Distribution index ; The iteration termination condition is: the evolution reaches 200 generations, or the relative change in the fitness value of the optimal individual is less than 10 over 30 consecutive generations. -6 At the same time, the top 5 individuals are retained in each generation.

[0020] Preferably, in step 4, the damage index The calculation formula is:

[0021] In the formula, This represents the inversion modulus of the driving lane in the corresponding structural layer. The inversion modulus of the hard shoulder at the same level. The value range is [0,1], and the damage index is... The closer the value is to 1, the more severe the damage at the corresponding layer.

[0022] Preferably, the pavement structure parameters include surface layer modulus, base layer modulus, subbase layer modulus, subgrade modulus, surface layer thickness, base layer thickness, and subbase layer thickness; The modulus of each structural layer of the pavement and the modulus and modulus parameters of each structural layer of the hard shoulder include the surface layer modulus, base layer modulus, subbase layer modulus and subgrade modulus.

[0023] The beneficial effects of this invention are as follows: By replacing the traditional theoretical model with a neural network surrogate model, the time required for a single forward modeling computation is reduced from seconds / minutes to milliseconds, significantly improving inversion efficiency and making rapid point-by-point inversion of large-scale road networks a reality. The global probabilistic search mechanism of the genetic algorithm effectively avoids getting trapped in local optima, demonstrating strong global optimization and multi-solution handling capabilities.

[0024] Using the hard shoulder as a natural benchmark, a structural layer damage index was defined, which transforms modulus decay into an intuitive percentage of damage, providing a direct and quantitative indicator for assessing pavement structure performance degradation, prioritizing maintenance, and designing reconstruction and expansion schemes.

[0025] Through multiple independent verification methods, including FWD verification, core sampling and on-site cross-section excavation, a complete technical closed loop is formed, from dynamic detection to intelligent inversion, damage assessment and on-site verification, ensuring the engineering applicability of the method and the reliability of the results. Attached Figure Description

[0026] Figure 1 The coefficient of determination (R²) for the neural network surrogate model in predicting the deflection slope at each measuring point (S10-S120). 2 Comparison chart of ( ) and root mean square error (RMSE); Figure 2 This is a convergence curve of the objective function error as a function of the number of iterations during the genetic algorithm inversion process. Figure 3 This is a comparison chart of the optimal inversion results for the driving lane and the measured TSD deflection slope. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments and accompanying drawings. This embodiment takes a section of the Jiangsu Riverside Expressway expansion project (chainage K42+290~K42+360) as the object. This section is a typical semi-rigid base asphalt pavement structure, with fatigue damage in the driving lanes and relatively intact hard shoulder structure.

[0028] This embodiment provides a method for pavement structure modulus inversion and damage evaluation based on deflection slope, and the specific steps are as follows.

[0029] Step 1: Data acquisition and preprocessing.

[0030] 1.1 Acquisition of raw data.

[0031] A high-speed deflectometer was used to test a semi-rigid base asphalt pavement. Multiple sets of laser Doppler vibration sensors were deployed longitudinally along the load center, with horizontal distances from the center of each sensor to the load center of 10 cm, 20 cm, 30 cm, 60 cm, 90 cm, and 120 cm, respectively. At a standard testing speed (60–80 km / h), the vehicle speed Vx (m / s) of the high-speed deflectometer was recorded in real time, and the pavement deformation velocity Vy (mm / s) at each laser vibration meter measuring point was obtained. Based on the formula S = Vy / Vx, the measured values ​​of the deflection slope for the six channels were calculated. , denoted as a vector The unit is 0.01 mm / m.

[0032] 1.2 Data quality control.

[0033] Outlier removal: Any channel or When the value is (0.01 mm / m), all data at that point are discarded.

[0034] Jump detection: When the difference between adjacent detection points in the same channel exceeds 3 times the historical standard deviation, it is considered an anomaly and is removed.

[0035] Missing value imputation: For individual missing values ​​caused by device packet loss, linear interpolation is used to imput them.

[0036] Step 2: Construct a neural network forward surrogate model.

[0037] 2.1 Training data generation and preprocessing.

[0038] Sample generation: Using layered elasticity theory program or finite element model, the Latin hypercube sampling (LHS) method is used to generate 1000 sets of pavement structure parameter combinations within the parameter range shown in Table 1, and the theoretical deflection slope at 20℃ corresponding to each set is calculated to form 1000 sets of "parameter-response" sample pairs.

[0039] Table 1: Range of Pavement Structure Parameters

[0040] Feature preprocessing: Input vector Take the natural logarithm of the modulus parameter. (i=1,2,3,4), and the three thickness parameters Keeping them unchanged, we then perform Z-score normalization on all seven features obtained after the transformation:

[0041]

[0042] in, It is the original value of the j-th feature. It is the standardized value of the j-th feature; These are the mean and standard deviation of the feature on the training set, respectively. The output variable (deflection slope) retains its original dimensions (0.01 mm / m) and is not standardized.

[0043] 2.2 Network Structure and Training.

[0044] Neural network structure: A multilayer perceptron (MLP) with 5 hidden layers is constructed. Each hidden layer is followed by a Swish activation function. The number of neurons is 128, 128, 256, 256, and 128 respectively. There is no activation function in the output layer.

[0045] The activation function takes the following form:

[0046]

[0047] In the formula: x represents the net input of the current neuron, which is dimensionless.

[0048] Loss function: Weighted root mean square error (RMSE) is used.

[0049]

[0050] Where N is the number of samples; This refers to the number of sensor channels; This represents the weight of the i-th channel, which is set to 1 by default. The deflection slope of the k-th sample and the i-th channel predicted by the neural network is expressed in units of 0.01 mm / m. The theoretical deflection slope for the k-th sample and the i-th channel is calculated, with units of 0.01 mm / m. The root mean square value of the theoretical deflection slope for all channels of the k-th sample is calculated using the following formula:

[0051]

[0052] Training configuration: The optimizer is Adam, the initial learning rate is 5e-4, which decreases by half every 200 epochs, the batch size is 32, and the maximum number of training epochs is 800. An early stopping mechanism is used (the training stops if the loss does not decrease for 25 consecutive epochs) to prevent overfitting.

[0053] Model performance: The overall coefficient of determination of the model on the validation set after training. The value is 0.9979, the overall RMSE is 0.01938 (0.01 mm / m), and the fitting accuracy of each channel is as follows: Figure 1 As shown, the surrogate model has high accuracy and good generalization ability.

[0054] Step 3: Modulus inversion optimization based on genetic algorithm.

[0055] 3.1 Definition of the inversion problem.

[0056] In this embodiment, the thickness parameters are taken as design values ​​(18cm for the surface layer, 40cm for the base layer, and 20cm for the subbase layer), and the inversion variables are four modulus parameters. For the measured deflection slope vector... (Unit: 0.01 mm / m) The goal is to find the optimal combination of moduli within the parameter space of Table 1. (Unit: MPa) The objective function is defined as the relative root mean square error between the measured deflection slope and the neural network prediction value. The objective function is minimized. :

[0057]

[0058] in, For the first The measured deflection slope of the channel is 0.01 mm / m. The predicted value (0.01 mm / m) is from the neural network surrogate model. This represents the number of sensor channels.

[0059] 3.2 Genetic Algorithm Design.

[0060] Encoding: Real number encoding, each individual is... .

[0061] Population: Size 1000, initialized using Latin hypercube sampling.

[0062] fitness: .

[0063] Selection: Tournament selection (Tournament size is 3).

[0064] Crossover: Simulated binary crossover (SBX), probability Distribution index .

[0065] Mutation: Polynomial mutation, probability Distribution index .

[0066] Elite strategy: Retain the top 5 individuals in each generation.

[0067] Termination conditions: Reaching the maximum number of generations 200, or the relative change in the optimal fitness value is less than 10 over 30 consecutive generations. -6 ,Right now

[0068]

[0069] In the formula, For the first The optimal fitness value of the generation population. It is a very small positive number ( ), used to avoid division by zero errors.

[0070] Convergence: such as Figure 2 As shown, the error decreases rapidly in the first few generations and stabilizes at around 0.073 after about 150 generations, indicating a robust convergence process.

[0071] 3.3 Fit quality and optimal solution.

[0072] Table 2 shows the optimal solution for modulus inversion. The results show that: the surface layer modulus is 8260 MPa, which is typical of the stiffness level of the superstructure of a semi-rigid base asphalt pavement; the base layer modulus reaches 14860 MPa, which is the stiffest layer in the overall structure, fully demonstrating the dominant role of the base layer in the deflection basin morphology; the subbase layer modulus is 5342 MPa, which is within a reasonable range and can form an effective load-bearing synergy with the higher base layer modulus; the subgrade modulus is 111 MPa, which conforms to the common mechanical characteristics of subgrades of expressways in Jiangsu Province. Figure 3 The comparison between the optimal deflection slope fitting curve obtained based on ANN-GA inversion and the measured TSD deflection slope is presented. Overall, the optimal modulus combination can well reproduce the measured TSD deflection slope, reflecting that the stiffness matching relationship between the various layers of the road section structure is reasonable, and the inversion results have good physical consistency.

[0073] Table 2: Optimal Solution for Modulus Inversion

[0074] Step 4: Pavement structure damage assessment.

[0075] Using the deflection slope data of the hard shoulder at the same station as input, the structural layer modulus of the hard shoulder in the same road section was inverted using the same method (step 3), and the results are shown in Table 3. Since the hard shoulder does not bear the main traffic load, its modulus can be regarded as the "initial health state" benchmark of the structure.

[0076] Table 3: Optimal Solution for Hard Shoulder Modulus Inversion

[0077] For critical load-bearing layers such as the base course and subbase course, the damage index is calculated using the following formula. :

[0078]

[0079] in, The inversion modulus of the driving lane (MPa) The inversion modulus of the hard shoulder is (MPa). The value ranges from [0,1], with larger values ​​indicating more severe damage.

[0080] By combining the solutions from Tables 2 and 3, the modulus damage of the driving lane base course is approximately 40%, and the modulus damage of the subbase course is approximately 57%. This quantitative damage index directly reflects the degree of performance degradation of the driving lane pavement structure layer under long-term load.

[0081] Step 5: Result verification.

[0082] To verify the reliability of this method, two verifications were conducted on the same road segment:

[0083] FWD Comparison and Verification: FWD detection was performed on 13 cross sections, and the modulus was inverted using the traditional iterative method. The relative deviations between the results and the modulus of each layer inverted by this method were all less than 10%, proving the accuracy of the results of this method.

[0084] Core sampling verification: Core samples were drilled at typical locations on the driving lane and hard shoulder, and indoor dynamic modulus tests were conducted. The measured base modulus ratio (driving lane / hard shoulder) was 0.45, which is in close agreement with the damage obtained by this method.

[0085] The above verification results show that the dynamic detection-intelligent inversion-damage assessment technology path proposed in this invention is reliable and effective, and can provide a complete system solution for rapid, non-destructive, and accurate evaluation of pavement structure performance.

Claims

1. A method for pavement structure modulus inversion and damage evaluation based on deflection slope, characterized in that, Includes the following steps: Step 1: Obtain the measured deflection slope data of multiple lanes of the road surface under traffic load, and preprocess the acquired data; Step 2: Use Latin hypercube sampling to generate multiple sets of pavement structure parameter combinations and calculate the theoretical deflection slope corresponding to each combination. Based on the pavement structure parameter combinations and the corresponding theoretical deflection slopes, construct and train a neural network forward surrogate model to predict the deflection slope. Step 3: With the goal of minimizing the relative root mean square error between the measured deflection slope and the predicted value of the neural network forward surrogate model, construct a fitness function and use a genetic algorithm to find the optimal solution for the inversion of the modulus of each structural layer of the pavement within the set parameter space. Step 4: Based on the measured deflection slope data of the hard shoulder in the same road segment, invert the modulus of each structural layer of the hard shoulder and use it as a health benchmark to calculate the damage index of the driving lane structural layer. .

2. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 1, characterized in that, In step 1, the measured deflection slope data includes multiple deflection slope values ​​at different horizontal distances from the load center.

3. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 1, characterized in that, Data preprocessing includes outlier removal, jump detection, and missing value imputation; among these, Outlier removal involves removing all data at the corresponding point when the deflection slope of any channel is less than 0 or greater than 200. Jump detection is used when the difference between adjacent detection points in the same channel exceeds 3 times the historical standard deviation, and the result is considered an anomaly and removed. Missing value imputation is used to fill in individual missing values ​​caused by device packet loss using linear interpolation.

4. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 1, characterized in that, In step 2, the specific configuration for constructing and training the neural network forward surrogate model is as follows: A multilayer perceptron network with 5 hidden layers was constructed, with each hidden layer followed by a Swish activation function. The number of neurons in each layer was configured as 128, 128, 256, 256, and 128, respectively, and no activation function was set for the output layer. The optimizer uses Adam and introduces an early stopping mechanism to prevent model overfitting.

5. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 4, characterized in that, In step 2, the input feature preprocessing method for the multilayer perceptron network is as follows: The network input vector contains modulus parameters and thickness parameters, and the natural logarithm of the modulus parameters of all inputs is taken. All processed features are Z-score standardized.

6. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 4 or 5, characterized in that, In step 2, the weighted root mean square error loss function used to train the neural network forward surrogate model is defined as: In the formula, N is the number of samples; This refers to the number of sensor channels; Let be the weight of the i-th channel; Let be the deflection slope of the k-th sample and the i-th channel predicted by the neural network; Calculate the theoretical deflection slope for the k-th sample and the i-th channel; Let the root mean square value of the theoretical deflection slope of all channels of the k-th sample satisfy:

7. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 1, characterized in that, In step 3, the objective function of the genetic algorithm for modulus optimization is defined as: In the formula, For the first Measured deflection slope of the channel The predicted value is from the neural network surrogate model. This represents the number of sensor channels.

8. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 7, characterized in that, In step 3, the specific running parameters of the genetic algorithm are configured as follows: Real number encoding was used, the population size was set to 1000, and population initialization was completed based on Latin hypercube sampling; The fitness function is defined as follows: ; Select the tournament selection with a scale of 3; The crossover operation is a simulated binary crossover, with crossover probability... Distribution index ; The mutation operation is a polynomial mutation, with a mutation probability of... Distribution index ; The iteration termination condition is that the evolution reaches 200 generations, or the relative change amount of the optimal individual fitness value is less than 10 for 30 generations in succession -6 Meanwhile, the top 5 optimal individuals are retained in each generation.

9. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 1, characterized in that, In step 4, the damage index The calculation formula is: In the formula, This represents the inversion modulus of the driving lane in the corresponding structural layer. The inversion modulus of the hard shoulder at the same level. The value range is [0,1], and the damage index is... The closer the value is to 1, the more severe the damage at the corresponding layer.

10. The method for pavement structure modulus inversion and damage evaluation based on deflection slope according to claim 5, characterized in that, The pavement structural parameters include the surface layer modulus, base layer modulus, subbase layer modulus, subgrade modulus, surface layer thickness, base layer thickness, and subbase layer thickness; the modulus of each structural layer of the pavement and the modulus of each structural layer of the hard shoulder, as well as the modulus parameters, all include the surface layer modulus, base layer modulus, subbase layer modulus, and subgrade modulus.

Images