A pavement service performance self-optimization prediction method

Abstract

This invention relates to a self-optimizing prediction method for pavement service performance, comprising: acquiring pavement vibration data, extracting vibration features, constructing a time-series feature tensor as model input, using the root mean square of vibration in the time domain as model output, and constructing a feature-decoupled multi-head deep network model; dividing the current pavement vibration data into a teacher stage and at least one incremental stage, calculating the comprehensive drift score of a single vibration feature, and calculating the drift score of a single sample based on this; training the model using the teacher stage data as a basic teacher model, and then using the incremental stage data for incremental learning to obtain a student model, the loss function of incremental learning being constructed in conjunction with the drift score of a single sample; outputting prediction results using the student model; when accumulating new data, using the student model as a new basic teacher model, and using the accumulated new data for incremental learning to achieve continuous model evolution. Compared with existing technologies, this invention has the advantages of achieving long-term and accurate predictions.

Description

Technical Field

[0001] This invention relates to the field of intelligent pavement sensing, and in particular to a method for self-optimizing prediction of pavement service performance. Background Technology

[0002] Pavement service performance refers to the comprehensive ability of a pavement to withstand traffic loads and resist environmental effects during its service life, reflecting the overall health status and functional maintenance level of the pavement. Whether it is airport pavement bearing the heavy loads of high-speed aircraft or road pavement bearing the long-term repeated rolling of vehicles, the cumulative effect of traffic loads will accelerate the degradation of pavement service performance, causing structural defects such as cracks, broken slabs, misalignment, and subsidence, as well as functional degradation such as decreased smoothness, reduced friction coefficient, and rutting. In severe cases, it will affect traffic safety and service levels. Therefore, accurately understanding the current status of pavement service performance and scientifically predicting its evolution trend is of great significance for the scientific maintenance, life extension, and full life cycle management of pavements.

[0003] In the initial research phase regarding pavement performance prediction, a research system dominated by mechanical-empirical models has been established both domestically and internationally. However, these mechanical-empirical models rely on deterministic formulas and have not fully considered the variability and uncertainty of loads, materials, and environment in actual use. Furthermore, they primarily focus on structural load-bearing capacity, while paying insufficient attention to the long-term evolution of functional performance such as smoothness and friction coefficient.

[0004] Subsequently, service performance prediction methods gradually shifted from deterministic models to probabilistic models. For example, the use of dynamic semi-Markov processes and Bayesian update methods can dynamically revise existing prediction models based on newly added test data, and gradually extend to the comprehensive evaluation of structural and functional performance, improving the reliability and accuracy of predictions. However, these methods still rely on low-frequency, discrete semi-manual testing methods (such as falling weight deflectometers, flatness meters, friction coefficient testing vehicles, etc.), resulting in insufficient data continuity, making it difficult to fully capture the long-term evolution of pavement service performance, and also incurring high testing costs and significant traffic disruption.

[0005] In recent years, with the development of sensing technology and artificial intelligence, data-driven methods such as deep learning have been gradually applied to pavement performance prediction. In the field of airport pavement, there are predictive studies targeting the response of rigid pavement structures, as well as predictive applications for the development of asphalt pavement distresses such as rutting and fatigue. In the field of road pavement, methods based on deep learning for pavement damage identification, smoothness prediction, and rutting prediction have been gradually applied. For example, Chinese patent CN116957136A discloses a pavement performance prediction method and device based on temporal deep learning. This method uses a recurrent neural network to build a prediction model, using historical maintenance behavior and pavement performance index data as input parameters. It combines a long short-term memory neural network or a gated recurrent unit and uses the Adam optimization algorithm for adaptive learning rate and parameter gradient updates to predict pavement performance.

[0006] However, most of these models are static prediction models trained on fixed datasets. As the pavement's service life extends, massive amounts of monitoring data reflecting the latest pavement condition accumulate. The evolutionary patterns contained in this new data may differ significantly from historical data (e.g., traffic volume growth, changes in environmental conditions, implementation of maintenance measures, etc.). Static models cannot continuously learn and evolve from new data, facing the challenge of catastrophic forgetting, causing the model's predictive performance to gradually decline over time, making it difficult to adapt to the long-term dynamic evolution trend of pavement service performance. Summary of the Invention

[0007] The purpose of this invention is to provide a self-optimizing prediction method for pavement service performance, which is aimed at long-term online evaluation and continuous optimization and updating of pavement service performance under complex operating conditions. This method achieves high-precision prediction of pavement vibration level and self-optimizing evolution of the model through a technical route of high-quality vibration feature construction, feature decoupling multi-head deep network and drift-aware incremental learning by teachers and students.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] A method for self-optimizing prediction of pavement service performance includes the following steps:

[0010] Acquire pavement vibration data under traffic loads and perform preprocessing;

[0011] Using the root mean square of vibration in the time domain as the target variable, feature engineering is performed on the preprocessed vibration data based on the target variable to extract vibration features;

[0012] Using vibration characteristics, traffic load, and time interval as input vectors at a single moment, and combining them with a time lag window to construct a time-series feature tensor as model input, and using the root mean square of vibration in the time domain as model output, a feature-decoupled multi-head deep network model is constructed.

[0013] The current pavement vibration data is divided into two key stages: the teacher stage and at least one incremental stage. The comprehensive drift score of a single vibration feature in each incremental stage is calculated by combining JS divergence and KS divergence, and the drift score of a single sample in each incremental stage is calculated based on the comprehensive drift score of a single vibration feature.

[0014] A feature-decoupled multi-head deep network model was trained using data from the teacher phase, serving as the base teacher model.

[0015] Based on the basic teacher model, incremental learning is performed using data from the incremental phase. If there are multiple incremental phases, incremental learning is performed multiple times in sequence to obtain the student model. The loss function of the incremental learning is the weighted sum of the new data supervision loss and the weighted distillation loss. The comprehensive weight in the weighted distillation loss is the product of the teacher confidence gate and the drift weight. The drift weight is calculated based on the drift score of a single sample in the current incremental phase.

[0016] Using the student model, the pavement service performance prediction results are output based on the real-time acquired vibration data;

[0017] New pavement vibration data is acquired in real time. When a certain amount of new data is accumulated, the student model obtained from the previous incremental learning is used as the new base teacher model, and the accumulated new data is used as the new incremental stage to carry out incremental learning again, so as to achieve continuous evolution of the model.

[0018] The feature engineering specifically involves: extracting candidate vibration features based on pavement vibration data, calculating the Pearson correlation coefficient between each candidate vibration feature and the target variable, and selecting features with a Pearson correlation coefficient greater than a preset threshold to form a vibration feature set.

[0019] The single-time input vector is represented as ,in, express The input vector at time t, express Traffic load at any given moment express The time interval of time, express Vibration characteristics at any given moment;

[0020] The temporal feature tensor is represented as follows: ,in, express The temporal feature tensor at time step, This represents the length of the time lag window.

[0021] The feature-decoupled multi-head deep network model includes the following sequentially connected components:

[0022] Feature extraction module: A two-layer one-dimensional convolutional neural network is used to extract local features from the input temporal feature tensor. Through convolution and pooling operations, the original input sequence is transformed into high-dimensional features containing spatiotemporal information.

[0023] Self-attention module: Introduces a multi-head self-attention mechanism to assign weights to the high-dimensional features extracted by the feature extraction module at each time step, adaptively highlighting key moments and key features that are more sensitive to service performance prediction, and obtaining a time-series feature sequence enhanced by attention;

[0024] Temporal modeling module: The module uses a long short-term memory neural network to model the temporal feature sequence after self-attention reweighting, and learns and memorizes the long-term dynamic evolution law and historical context dependency relationship in the pavement service performance degradation process;

[0025] Regression prediction module: The representation vectors obtained by the time series modeling module for each feature channel are concatenated and input into the fully connected regression network. The learned high-dimensional abstract representation is mapped into specific numerical output as the prediction result of the root mean square of vibration in the time domain.

[0026] The loss functions used when training feature-decoupled multi-head deep network models using data from the teacher phase include Huber loss and trend consistency loss.

[0027] The comprehensive drift score for calculating individual vibration features in each incremental stage by combining JS divergence and KS divergence is as follows:

[0028] Let the current incremental stage be denoted as Then the first One characteristic in the teaching stage T and the current incremental phase and all previous incremental phases The empirical distributions on are respectively , ≥1 corresponds to any pair of stages Define vibration characteristics In the stage Drift intensity for:

[0029] ;

[0030] in, For the first The KS divergence of a vibration feature between stages A and B For the first The JS divergence of a vibration feature between stages A and B;

[0031] Take the maximum value of the drift intensity of all phase pairs corresponding to the current increment phase to obtain the first drift intensity in the current increment phase. A comprehensive drift score for each vibration characteristic:

[0032] ;

[0033] in, For the current incremental stage Vibration characteristics The overall drift score.

[0034] The comprehensive drift score calculation based on a single vibration feature for each incremental stage specifically involves:

[0035] ;

[0036] ;

[0037] in, The total dimension of vibration characteristics. For the current incremental stage Vibration characteristics The overall drift score, For the current incremental stage The average drift level of all vibration characteristics within the body. For the current incremental stage medium sample Drift rating , , Representing samples respectively The timestamp, the start time and end time of the current incremental phase.

[0038] The new data supervision loss is expressed as:

[0039] ;

[0040] in, Loss due to new data supervision The size of the current training batch. For the current incremental stage sample Real data, For the student model's prediction results on the new data, Here is the Huber loss function.

[0041] The weighted distillation loss is expressed as:

[0042] ;

[0043] in, For weighted distillation losses, The predicted value for the teacher model. For the student model's prediction results on the new data, The distillation temperature. The size of the current training batch. For the sample The overall weight, , For teachers, confidence gating For drift weights;

[0044] The drift weight is defined as:

[0045] ;

[0046] in, The drift sensitivity coefficient, For the current incremental stage medium sample Drift rating;

[0047] The teacher confidence gating is defined as follows:

[0048] ;

[0049] in, For the current incremental stage sample Real data, The residual quantiles of the teachers on the incremental training set.

[0050] The loss function for incremental learning is expressed as:

[0051] ;

[0052] in: , representing the weight, which gradually increases with the number of training rounds. Loss due to new data supervision This represents the weighted distillation loss.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] (1) This invention can accurately predict the service performance index of the pavement under the next traffic load, and evaluate the service performance status of the pavement in real time and predict the evolution trend of the service performance of the pavement. It continuously learns from the new monitoring data and adaptively predicts the evolution of the pavement performance over time. It effectively overcomes the problem of performance decay caused by changes in data distribution in traditional static artificial intelligence models, and ensures the accuracy and reliability of the prediction model throughout the entire life cycle, providing technical support for forward-looking maintenance decisions of the pavement.

[0055] (2) The present invention introduces the combined calculation of JS divergence and KS divergence, which quantifies the difference in feature distribution between new data and old data, and can capture data drift more comprehensively. By calculating the comprehensive drift score, the model can identify which vibration features have changed significantly, so that these changes can be focused on in subsequent learning, avoiding the model disorder that may be caused by blindly learning all new data.

[0056] (3) This invention employs a teacher-student distillation architecture and a weighted loss function, enabling the model to selectively retain old knowledge while learning new data. This prevents the catastrophic forgetting problem of completely forgetting historical patterns due to learning new features, and also achieves low-cost continuous evolution of the model, adapting to the dynamic changes throughout the entire life cycle of the pavement. Among them, the comprehensive weight in the weighted distillation loss of the loss function is the product of teacher confidence gating and drift weight. Teacher confidence gating dynamically adjusts the credibility of the teacher model in the distillation process, preventing the teacher model from giving incorrect guidance on highly variable data. Drift weight is calculated based on the drift score of a single sample in the current incremental stage, which can assign higher learning weights to new samples that differ greatly from the historical distribution. For data that has drifted, the model will focus on learning with high drift weights; for parts that do not seriously conflict with historical knowledge, the model focuses on maintaining stability through gating balance. This effectively prevents the model from overfitting on new data and underfitting on old knowledge.

[0057] (4) The service environment of airports and road pavements is time-varying. This invention can capture these dynamic changes in real time through drift scoring, quickly adapt to the new state in the incremental learning stage, and retain the underlying knowledge of long-cycle fatigue damage to ensure that the prediction model can accurately reflect the new performance degradation rate. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method of the present invention;

[0059] Figure 2 This is a schematic diagram of the feature-decoupled multi-head deep network model structure of the present invention;

[0060] Figure 3 This is a schematic diagram of the first incremental learning result in one embodiment of the present invention;

[0061] Figure 4 This is a schematic diagram of the second incremental learning result in one embodiment of the present invention. Detailed Implementation

[0062] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0063] This embodiment provides a self-optimizing prediction method for pavement service performance. It collects high-frequency, continuous vibration responses induced by traffic loads using distributed optical fibers embedded in the pavement. A deep learning-based service performance prediction model is constructed using multi-dimensional vibration characteristic data, applicable to airport and road pavements. Furthermore, an incremental learning mechanism based on long-term monitoring data is introduced, enabling the model to continuously learn and update itself online using new data without forgetting old knowledge. This avoids problems such as excessive computational resource consumption, loss of historical knowledge, and model service interruption caused by repeated model training, thereby improving the long-term adaptability and engineering practical value of the prediction model.

[0064] like Figure 1 As shown, the method includes the following steps:

[0065] S1: Obtain pavement vibration data under traffic load and perform preprocessing.

[0066] This embodiment collects high-frequency, continuous vibration responses induced by traffic loads using distributed optical fibers embedded in the pavement, and uses the interquartile range (IQR) method to identify and remove outliers; the moving average method is used to smooth the data, suppress noise interference, and improve data quality.

[0067] S2, using the root mean square of vibration in the time domain as the target variable, perform feature engineering on the preprocessed vibration data based on the target variable to extract vibration features.

[0068] Based on literature review and previous research, this embodiment identifies the root mean square (RMS) of vibration in the time domain as a key indicator characterizing the pavement vibration level, and uses it as the target variable for model prediction. The calculation formula is as follows:

[0069] ;

[0070] in, For the first Vibration signals at each sampling point (time). This indicates the total number of sampling points.

[0071] In subsequent steps, all feature engineering, model building, and evaluation use RMS as the sole prediction target to ensure methodological consistency.

[0072] Subsequently, candidate vibration features were extracted based on pavement vibration data, the Pearson correlation coefficient between each candidate vibration feature and the target variable was calculated, and features with a Pearson correlation coefficient greater than 0.6 were selected to form a vibration feature set.

[0073] S3 uses vibration characteristics, traffic load, and time interval as input vectors at a single moment, and combines them with a time lag window to construct a time-series feature tensor as the model input. The vibration time domain root mean square is used as the model output to construct a feature-decoupled multi-head deep network model.

[0074] In this embodiment, to address the temporal irregularities of data collection, the time interval between adjacent data points is used as a one-dimensional key feature to effectively capture the true temporal pattern of load action. At the same time, different traffic loads are also embedded in the feature space, enabling the model to perceive the true temporal pattern of load action.

[0075] Therefore, the input vector at a single time moment is represented as ,in, express The input vector at time t, express Traffic load at any given moment express The time interval of time, express The vibration characteristics at time t, by It consists of vibration characteristic values ​​in several dimensions.

[0076] To extract the temporal dependence of vibration features, a time lag window is constructed for each feature dimension, the window being defined by the current time step. Previous consecutive The feature values ​​at each time step constitute the input vector. :

[0077] ;

[0078] in, express The temporal feature tensor at time step, This represents the length of the time lag window.

[0079] In this embodiment, the feature-decoupled multi-head deep network model adopts a parallel multi-head network architecture based on feature decoupling. Unlike traditional methods that mix all features in the input, this architecture constructs an independent processing channel for each feature dimension, avoiding forced sharing of convolution / gating parameters and reducing mutual interference between features.

[0080] like Figure 2 As shown, the feature-decoupled multi-head deep network model includes the following sequentially connected components:

[0081] Feature extraction module: Employs a two-layer one-dimensional convolutional neural network (CNN) to extract the input temporal feature tensor. Local feature extraction is performed, and the original input sequence is transformed into high-dimensional features containing spatiotemporal information through convolution and pooling operations;

[0082] Self-attention module: After convolutional feature extraction, a multi-head self-attention mechanism is introduced to weight the high-dimensional features extracted by the feature extraction module at each time step, adaptively highlighting the key moments and key features that are more sensitive to service performance prediction, suppressing noise and redundant information, and obtaining a time-series feature sequence enhanced by attention.

[0083] Temporal modeling module: Long Short-Term Memory Neural Network (LSTM) is used to model the temporal feature sequence after self-attention reweighting, learn and memorize the long-term dynamic evolution law and historical context dependency relationship in the pavement service performance degradation process;

[0084] Regression prediction module: The representation vectors obtained by the time series modeling module for each feature channel are concatenated and input into the fully connected regression network. The learned high-dimensional abstract representation is mapped into a specific numerical output, which serves as the prediction result of the vibration time domain root mean square (RMS).

[0085] S4 divides the current pavement vibration data into two key stages: the teacher stage and at least one incremental stage. Combines JS divergence and KS divergence to calculate the comprehensive drift score of a single vibration feature in each incremental stage, and calculates the drift score of a single sample in each incremental stage based on the comprehensive drift score of a single vibration feature.

[0086] During the long-term service of pavement, the statistical distribution of vibration characteristics changes significantly due to the continuous evolution of load patterns, environmental conditions, and structural performance, leading to a decline in the performance of predictive models trained on historical data. This phenomenon is known as "data drift" in machine learning. If not identified and addressed, it will cause a continuous decline in the model's predictive ability, i.e., the "model aging" problem. To quantitatively assess the degree of change in data distribution across different time periods, Jensen-Shannon (JS) divergence and Kolmogorov–Smirnov (KS) divergence were used as drift metrics to systematically analyze the differences in characteristic distributions between several key time periods.

[0087] Let the current incremental stage be denoted as Then the first One characteristic in the teaching stage T and the current incremental phase and all previous incremental phases The empirical distributions on are respectively , ≥1 corresponds to any pair of stages Define vibration characteristics In the stage Drift intensity for:

[0088] ;

[0089] in, For the first The KS divergence of a vibration feature between stages A and B For the first The JS divergence of a vibration feature between stages A and B;

[0090] Take the maximum value of the drift intensity of all phase pairs corresponding to the current increment phase to obtain the first drift intensity in the current increment phase. A comprehensive drift score for each vibration characteristic:

[0091] ;

[0092] in, For the current incremental stage Vibration characteristics The overall drift score. The larger the value, the more significant the distributional differences of this feature across at least one pair of time periods in multiple stages.

[0093] Subsequently, the drift score of an individual sample in each incremental stage is calculated based on the comprehensive drift score of a single vibration feature:

[0094] ;

[0095] ;

[0096] in, The total dimension of vibration characteristics. For the current incremental stage Vibration characteristics The overall drift score, For the current incremental stage The average drift level of all vibration characteristics within the body. For the current incremental stage medium sample Drift rating , , Representing samples respectively The timestamp, the start time and end time of the current incremental phase.

[0097] This embodiment uses vibration data from November 29, 2020 to August 24, 2025 as an example for detailed explanation. First, the stages are divided, including: the teacher stage. T (November 29, 2020 to March 31, 2021) First Incremental Phase (March 31, 2021 to August 21, 2024), Second Incremental Phase (August 21, 2024 to August 24, 2025), which means there are a total of 2 incremental phases.

[0098] For any two stages and For each feature dimension, kernel density estimation (KDE) is first performed to obtain a continuous probability density function, and then JS divergence is used as the first drift measure.

[0099] ;

[0100] ;

[0101] in, , These are the characteristic probability distributions for the two stages; It is the middle distribution between the two distributions; for Divergence is used to measure the difference between two probability distributions. The divergence value ranges from [0,1], and the larger the value, the more significant the distribution difference.

[0102] Based on this, a combined drift strength index of the Kolmogorov–Smirnov (KS) statistic and JS divergence is introduced. Let the... These characteristics are present in the teacher stage ( T ), first incremental stage ( ) and the second incremental stage ( The empirical distributions on ) are respectively , corresponding to any pair of stages The drift intensity of this feature on this time pair is defined as:

[0103] ;

[0104] in, For the first The KS statistic of a feature between stages A and B This represents the corresponding JS divergence.

[0105] Furthermore, the maximum value is taken for the drift intensity of the time pair corresponding to the current incremental stage, to obtain the first incremental stage. The comprehensive drift score of each feature, where

[0106] ;

[0107] ;

[0108] in, , The first incremental stage Second Incremental Phase Vibration characteristics The overall drift score.

[0109] Subsequently, the drift score of a single sample in each incremental stage is calculated based on the comprehensive drift score of a single vibration feature. Taking the second incremental stage as an example, the calculation method is as follows:

[0110] ;

[0111] ;

[0112] in, For the second incremental stage The average drift level of all vibration characteristics within the body. For the second incremental stage medium sample Drift rating , , Representing samples respectively The timestamp, the start time of the current incremental phase (August 21, 2024), and the end time (August 24, 2025).

[0113] S5 utilizes data from the teacher phase to train a feature-decoupled multi-head deep network model, which serves as the base teacher model.

[0114] In this embodiment, to balance numerical accuracy and consistency of time trends, the loss function used when training features from the teacher phase to decouple the multi-head deep network model includes Huber loss and trend consistency loss.

[0115] The single-sample Huber loss is defined as:

[0116] ;

[0117] in, For the sample The true value, For the sample The model's predicted value.

[0118] When the batch size is B, take the average: .

[0119] Suppose the samples in the current batch are arranged in chronological order, and the actual values ​​and predicted values ​​are respectively... , Then the adjacent differences are:

[0120] .

[0121] Trend consistency loss is defined as:

[0122] ;

[0123] like and Same sign (actual and predicted trends are consistent). The penalty is 0.

[0124] If the two signs are different (the trends are opposite) The punishment is ;

[0125] The total loss for each batch in the basic teacher model is:

[0126] ;

[0127] in, The trend loss weighting balances point value precision (Huber) with trend consistency.

[0128] In this embodiment, the Adam optimization algorithm is used for training, combined with weight decay (L2 regularization) and gradient pruning to control model complexity and training stability. An early stopping strategy is used to automatically terminate training based on the validation set loss to avoid overfitting.

[0129] S6, based on the basic teacher model, uses incremental learning with data from the incremental phase. If there are multiple incremental phases, incremental learning is performed multiple times in sequence to obtain the student model.

[0130] The model trained on historical data is solidified as the teacher model, and its parameters are frozen as a source of prior knowledge; the student model inherits its architecture and parameters as the starting point for incremental learning.

[0131] The loss function for incremental learning is a weighted sum of the new data supervision loss and the weighted distillation loss:

[0132] ;

[0133] in: , representing the weight, which gradually increases with the number of training rounds. Loss due to new data supervision This represents the weighted distillation loss.

[0134] The new data supervision loss is expressed as:

[0135] ;

[0136] in, The size of the current training batch. For the current incremental stage sample Real data, For the student model's prediction results on the new data, Here is the Huber loss function.

[0137] Weighted distillation loss is expressed as:

[0138] ;

[0139] in, The predicted value for the teacher model. The distillation temperature. For the sample The overall weight, , For teachers, confidence gating This is the drift weight.

[0140] The drift weight is defined as:

[0141] ;

[0142] in, The drift sensitivity coefficient, For the current incremental stage medium sample Drift rating. When When the sample size is small (close to the old distribution), Distillation has a strong effect, which is beneficial for preserving old knowledge;

[0143] when When the sample size is large (far from the old distribution), As the distillation effect decreases or even approaches zero, the student model can learn new patterns more fully.

[0144] To prevent the imposition of strong distillation constraints on students when teachers' predictions are clearly inaccurate, this invention introduces a residual-based teacher confidence gating:

[0145] ;

[0146] in, This represents the residual quantiles of the teachers on the incremental training set. When the teacher error is small, Smaller This indicates that the teacher's output can be trusted; however, if the teacher's error is significant, The distillation term for the corresponding sample is significantly weakened or approximately turned off, retaining only the supervised loss based on the true label, thus avoiding the student model inheriting obviously erroneous old knowledge.

[0147] The student model minimizes simultaneously through backpropagation. and We used a reduced learning rate to fine-tune the parameters and ensure model stability.

[0148] S7 uses a student model to output pavement service performance prediction results based on real-time acquired vibration data.

[0149] S8 acquires new pavement vibration data in real time. When a certain amount of new data is accumulated, the student model obtained from the previous incremental learning is used as the new basic teacher model, and the accumulated new data is used as the new incremental stage to carry out incremental learning again, so as to achieve continuous evolution of the model.

[0150] For example, after accumulating enough data from August 25, 2025 to July 30, 2026, this period will be used as the third incremental phase. The student model trained in the second incremental phase will be used as the new base teacher model for incremental learning again.

[0151] This embodiment takes the pavement of a hub airport in China as the application scenario. Based on the vibration characteristic data from November 2020 to August 2025, a total of 2,541 records were collected, with each record corresponding to one aircraft taxiing event.

[0152] IQR outlier cleaning and sliding window smoothing (window size 5) were performed on the raw vibration characteristic data to fill in missing machine weight values.

[0153] Using RMS as the performance indicator for pavement service, correlation analysis was used to select 10-dimensional vibration features, including root mean square (RMS), mean square value, peak value, valley value, root mean square amplitude, margin index, mean value, dominant frequency, center of gravity frequency, and spectral entropy in the time domain. These features were then combined with aircraft load and time difference characteristics to form a 12-dimensional input feature set. To verify the effectiveness of the algorithm, the data was divided according to Table 1.

[0154] Table 1 Dataset Partitioning

[0155]

[0156] (1) Teacher Model

[0157] To simulate real-world engineering applications and evaluate the model's temporal generalization ability, the dataset was divided in strict chronological order: the baseline teacher model training data spanned from November 29, 2020 to March 31, 2021, totaling 1545 flights, and was divided into training, validation, and test sets in a 7:2:1 ratio. The prediction results are shown in the table below:

[0158] Table 2 Teacher Model Evaluation Indicators

[0159]

[0160] (2) Data drift detection

[0161] Multi-stage drift detection was performed on all input features, specifically divided into three key time periods: the teacher stage. T (November 29, 2020 to March 31, 2021) First Incremental Phase (March 31, 2021 to August 21, 2024), Second Incremental Phase (From August 21, 2024 to August 24, 2025), the test results are shown in Table 3.

[0162] Table 3. Results of Multi-Stage Data Drift Detection

[0163]

[0164] (3) First incremental learning

[0165] The first incremental learning employs a knowledge distillation framework. The student model learns new knowledge while retaining old knowledge by simultaneously optimizing the supervised loss and distillation loss. The hyperparameters are configured as follows: distillation temperature T = 3.0, tradeoff coefficient α = 0.5, and learning rate lr = 1 × 10⁻⁶. -4 The batch size was 32, the maximum number of training epochs was 300, and the early stopping patience was 80. Early stopping was triggered in the 127th epoch during actual training. The adaptability of the teacher model and the student model to new data was as follows: Figure 3 As shown in the figure, student model 1 is the student model obtained after the first incremental learning. The figure shows that student model 1 effectively adapted to the data distribution of the new time period through incremental learning, thus improving its generalization performance.

[0166] (4) Second Incremental Learning

[0167] Because the second incremental learning dataset has less data and a more significant data distribution shift, the hyperparameters were further adjusted as follows: the distillation temperature was reduced to T=2.5 to make the soft labels closer to the hard labels, accelerating adaptation to the new distribution; the weighting coefficient was reduced to α=0.4 to give greater weight to the distillation loss, strengthening knowledge preservation and avoiding overfitting on small samples; the learning rate was adjusted to lr=8×10⁻⁶. -5 The step size is updated with more refined parameters; the batch size is increased to 64, improving the stability of gradient estimation.

[0168] Figure 4 The comparison of prediction results from the second incremental learning is shown. Figure 4 (a) and Figure 4 (b) represents student model 1 (the student model obtained after the first incremental learning) and student model 2 (the student model obtained after the second incremental learning), respectively. Performance on the stage test set. The results show that, although The sample size was smaller (only 246 flights) and the data drift was more significant, but Student Model 2 effectively captured new evolutionary patterns through a self-optimizing prediction method, and its prediction performance was better than that of Student Model 1 in the previous stage.

[0169] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for self-optimizing prediction of pavement service performance, characterized in that, Includes the following steps: Acquire pavement vibration data under traffic loads and perform preprocessing; Using the root mean square of vibration in the time domain as the target variable, feature engineering is performed on the preprocessed vibration data based on the target variable to extract vibration features; Using vibration characteristics, traffic load, and time interval as input vectors at a single moment, and combining them with a time lag window to construct a time-series feature tensor as model input, and using the root mean square of vibration in the time domain as model output, a feature-decoupled multi-head deep network model is constructed. The current pavement vibration data is divided into two key stages: the teacher stage and at least one incremental stage. The comprehensive drift score of a single vibration feature in each incremental stage is calculated by combining JS divergence and KS divergence, and the drift score of a single sample in each incremental stage is calculated based on the comprehensive drift score of a single vibration feature. A feature-decoupled multi-head deep network model was trained using data from the teacher phase, serving as the base teacher model. Based on the basic teacher model, incremental learning is performed using data from the incremental phase. If there are multiple incremental phases, incremental learning is performed multiple times in sequence to obtain the student model. The loss function of the incremental learning is the weighted sum of the new data supervision loss and the weighted distillation loss. The comprehensive weight in the weighted distillation loss is the product of the teacher confidence gate and the drift weight. The drift weight is calculated based on the drift score of a single sample in the current incremental phase. Using the student model, the pavement service performance prediction results are output based on the real-time acquired vibration data; New pavement vibration data is acquired in real time. When a certain amount of new data is accumulated, the student model obtained from the previous incremental learning is used as the new base teacher model, and the accumulated new data is used as the new incremental stage to carry out incremental learning again, so as to achieve continuous evolution of the model.

2. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The feature engineering specifically involves: extracting candidate vibration features based on pavement vibration data, calculating the Pearson correlation coefficient between each candidate vibration feature and the target variable, and selecting features with a Pearson correlation coefficient greater than a preset threshold to form a vibration feature set.

3. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The single-time input vector is represented as ,in, express The input vector at time t, express Traffic load at any given moment express The time interval of time, express Vibration characteristics at any given moment; The temporal feature tensor is represented as follows: ,in, express The temporal feature tensor at time step, This represents the length of the time lag window.

4. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The feature-decoupled multi-head deep network model includes the following sequentially connected components: Feature extraction module: A two-layer one-dimensional convolutional neural network is used to extract local features from the input temporal feature tensor. Through convolution and pooling operations, the original input sequence is transformed into high-dimensional features containing spatiotemporal information. Self-attention module: Introduces a multi-head self-attention mechanism to assign weights to the high-dimensional features extracted by the feature extraction module at each time step, adaptively highlighting key moments and key features that are more sensitive to service performance prediction, and obtaining a time-series feature sequence enhanced by attention; Temporal modeling module: The module uses a long short-term memory neural network to model the temporal feature sequence after self-attention reweighting, and learns and memorizes the long-term dynamic evolution law and historical context dependency relationship in the pavement service performance degradation process; Regression prediction module: The representation vectors obtained by the time series modeling module for each feature channel are concatenated and input into the fully connected regression network. The learned high-dimensional abstract representation is mapped into specific numerical output as the prediction result of the root mean square of vibration in the time domain.

5. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The loss functions used when training feature-decoupled multi-head deep network models using data from the teacher phase include Huber loss and trend consistency loss.

6. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The comprehensive drift score for calculating individual vibration features in each incremental stage by combining JS divergence and KS divergence is as follows: Let the current incremental stage be denoted as Then the first One characteristic in the teaching stage T and the current incremental phase and all previous incremental phases The empirical distributions on are respectively , ≥1 corresponds to any pair of stages Define vibration characteristics In the stage Drift intensity for: ； in, For the first The KS divergence of a vibration feature between stages A and B For the first The JS divergence of a vibration feature between stages A and B; Take the maximum value of the drift intensity of all phase pairs corresponding to the current increment phase to obtain the first drift intensity in the current increment phase. A comprehensive drift score for each vibration characteristic: ； in, For the current incremental stage Vibration characteristics The overall drift score.

7. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The comprehensive drift score calculation based on a single vibration feature for each incremental stage specifically involves: ； ； in, The total dimension of vibration characteristics. For the current incremental stage Vibration characteristics The overall drift score, For the current incremental stage The average drift level of all vibration characteristics within the body. For the current incremental stage medium sample Drift rating , , Representing samples respectively The timestamp, the start time and end time of the current incremental phase.

8. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The new data supervision loss is expressed as: ； in, Loss due to new data supervision The size of the current training batch. For the current incremental stage sample Real data, For the student model's prediction results on the new data, Here is the Huber loss function.

9. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The weighted distillation loss is expressed as: ； in, For weighted distillation losses, The predicted values ​​are from the teacher model. For the student model's prediction results on the new data, The distillation temperature. The size of the current training batch. For the sample The overall weight, , For teachers, confidence gating For drift weights; The drift weight is defined as: ； in, The drift sensitivity coefficient, For the current incremental stage medium sample Drift rating; The teacher confidence gating is defined as follows: ； in, For the current incremental stage sample Real data, The residual quantiles of the teachers on the incremental training set.

10. The method for self-optimizing prediction of pavement service performance according to claim 1, characterized in that, The loss function for incremental learning is expressed as: ； in: , representing the weight, which gradually increases with the number of training rounds. Loss due to new data supervision This represents the weighted distillation loss.

Images