A method for predicting rutting on asphalt pavement based on multivariate transfer entropy and graph neural network
Through the method based on multivariate transfer entropy and graph neural network, the problems of high cost and low efficiency in rut prediction are solved, more accurate and efficient rut prediction are achieved, and the basis for asphalt pavement design and maintenance is improved.
Patent Information
- Application Number
- CN202210635944.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-07
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-06-07
AI Technical Summary
The prior art has problems of high cost and low efficiency in rut prediction, and it is difficult to accurately predict the development curve and influencing factors of asphalt pavement ruts.
Using a multivariate transfer entropy and graph neural network method, the time series of ruts and influencing factors is converted into graph structures, node embedding is used using GRU and self-attention mechanism, and multi-step prediction is performed using the spectrum-based graph neural network model. , calculate the transfer entropy to measure the causal relationship and reduce the impact of redundant variables.
It improves the accuracy and generalization ability of rut prediction, reduces the computational complexity, and can more accurately predict the development trend and key influencing factors of ruts.
Smart Images

Figure CN115048862B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a prediction method, in particular to an asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural network, belonging to the technical field of asphalt pavement performance prediction. Background Art
[0002] Rutting deformation is one of the major defects of high-grade asphalt pavements. It not only reduces the comfort and durability of high-speed vehicles but also poses potential safety hazards to vehicles during driving. Therefore, rutting prediction is a crucial task in the field of pavement structures. Accurately predicting asphalt pavement rutting is also an important basis for guiding asphalt pavement design, maintenance, and repair. To obtain accurate rutting development curves and establish the relationship between rutting and influencing factors, extensive research has been conducted on rutting test methods over the past few decades. Rutting test methods can be divided into laboratory rutting tests, accelerated pavement tests (APTs), and long-term pavement performance monitoring (LTPPs). Full-scale tracks, as a typical APT method, can be used to simultaneously compare the temporal and spatial equivalence of multiple pavement structures and material properties, and have attracted the attention of many research institutions and researchers. Consequently, many countries have constructed full-scale tracks, such as the AASHTO, MnROAD, and NCAT tracks in the United States, the IFSTTAR test track in France, and the HVS track in South Africa. The Research Institute of Track (RIOHTrack) of the Ministry of Highways and Transport was established in Beijing, China in 2017 to research and establish a next-generation long-life pavement design and construction technology system in China. The track consists of multiple sets of pavement structures and a large amount of field data was collected. Summary of the Invention
[0003] This invention addresses the challenges of the existing technology by providing an asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural networks. Considering the high cost of collecting experimental data for asphalt pavement rutting, the invention provides an asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural networks. The method consists of three modules: a graph representation module based on multivariate transfer entropy (mTE), a node embedding module based on GRU-attention, and a rutting prediction model based on GNN.
[0004] In order to achieve the above objectives, the specific technical solutions of the present invention are as follows:
[0005] Step 1: Collect historical asphalt pavement service performance monitoring data and asphalt pavement performance influencing factors data, and classify according to the collected data;
[0006] Step 2: Convert the time series of rutting and multiple variables affecting rutting into a graph structure. The number of nodes in the graph is the number of variables, and the edges are defined as the causal relationship between the typical variables and the variables to be predicted, which is obtained by calculating the transfer entropy between the variables.
[0007] Step 3: Embed the nodes in the graph into a GRU and self-attention mechanism. This can be considered a pre-trained model.
[0008] Step 4: Use a spectrum-based graph neural network model to create a multi-layer model structure to train the variable node relationship graph obtained in the above steps and output a multi-step prediction sequence.
[0009] As an improvement of the present invention, step 1 is as follows: collect the monitoring data of the service performance of asphalt pavement over the years and the data of the factors affecting the performance of asphalt pavement, classify the collected data, determine the input and output variables of the rutting prediction model, and use the time series of the road rutting, deflection center value, deflection basin area, temperature, load and other data as the input of the model. The input time series I∈R n×t , where n is the number of rutting and other related variables, t is the size of each time series, and the multi-step prediction data about the road rutting data is output to facilitate the calculation of the transfer entropy of each variable to the predicted variable in step 2.
[0010] As an improvement of the present invention, step 2 is as follows: the variable n in step 1 is represented by a graph structure G = {V, E}, where V = {v1, v2, ...}, E = {e 1,2 ,e 1,3 ,…}, respectively represent the relationship between variables, where the edge relationship is obtained by calculating the multivariate transfer entropy. Transfer entropy is a non-parametric statistical method for measuring causal relationships between processes based on information theory and ordered Markov processes. Based on information entropy and calculated from conditional entropy, information entropy is used to measure the average level of uncertainty in a process. Given a process X t =x1,x2,…,x t , its information entropy is expressed as
[0011]
[0012] where p(x i ) is the variable component x i The probability of the process is greater, the greater the amount of information in the process, the greater the information entropy of the process. Conditional entropy is a measure of uncertainty in process Y given another process X. Consider two processes X t =x1,x2,…,x t and Y t=y1,y2,…,y t Represent the historical values of time series X and Y respectively. Then the conditional entropy of given X and Y can be expressed as
[0013]
[0014] Then the mutual information from process X to Y can be expressed as
[0015]
[0016] The reduction in the mean uncertainty of Y is quantified.
[0017] X t =x τ-1 ,x τ-2 ,…,x τ-t and Y t =y τ-1 ,y τ-2 ,…,y τ-t As the characteristic process of X and Y, then from X t to Y t The transfer entropy can be expressed as
[0018]
[0019] in is the predicted value y τ and The conditional entropy of the joint distribution, It is y τ and historical sequence The conditional entropy of , k is the size of the sliding window. The transfer entropy is defined as τ ,VARIABLE X t and historical variable Y t The amount of information reduction, the value of transfer entropy has been shown to be equal to the value of conditional mutual information.
[0020] For a multivariate process X=X1,X2,…,X n The collective transfer entropy of Y is defined as the historical information of Y, each element subset Z (Z∈X n ) and the sum of other uncertainties:
[0021]
[0022] in It is an active information store used to capture historical information about Y; is a joint vector with a sliding window size of k; is the sum of the conditional mutual information increments, where each term is from the subset of accidental sources Z <n With information about X and Y, these subsets can be defined as U Y The inherent uncertainty of Y is due to the collective transfer entropy. Multivariate transfer entropy is introduced to capture causal relationships that manifest only through interactions, thereby reducing redundant links. Furthermore, a threshold can be set, setting transfer entropies below the threshold to zero. This reduces the influence of low-correlation variables and improves the method's generalization ability.
[0023] As an improvement of the present invention, step 3 is as follows: GRU is used to extract the time features of each input variable, which is regarded as the representation of the node embedding multivariate transfer entropy graph. The hidden unit of GRU is driven by LSTM unit, and the dependency relationship of different time series is captured through the gate mechanism. Given an input time series X = x1, x2, ..., x t and the previous hidden state H = h1,h2,…,h t-1 , then the hidden state h t At time step t it is defined as
[0024] h t =Z t ⊙tanh(W h x t +U h (r t ⊙h t-1 )+(1-Z t ⊙h t-1 ))
[0025] z t =σ(W z x t +U z h t-1 ), r t =σ(W r x t +U r h t-1 )
[0026] where z t and r t are the update gate and reset gate respectively; σ and tanh are the logical sigmoid function and the tangent hyperbolic function; ⊙ refers to the multiplication of the corresponding elements of the two matrices; W h 、U h 、W z 、U z 、W r 、U r They are all learnable parameters. Update gate z tThe output range is [0,1]. The closer it is to 1, the more information is remembered. Then use the attention-GRU based encoder module to generate the input time series represented as emb v ,v∈G is used as a node embedded in the multivariate transfer entropy graph. The output of the encoder at step n is
[0027]
[0028]
[0029] The vectors q, k, v are vector representations of the time steps in the intermediate matrices Q, K, V, which refer to the linear transformation of the input time series, i.e. Q = W Q X,K=W K X,V=W K X;α(q i ,k j ) refers to q i ,k j The similarity at time steps i and j is calculated using the mean squared error (MSE) function for model pre-training. Transfer entropy is used to measure the impact of a variable's historical data on the predicted variable's rutting data, which helps identify key variables and reduce the impact of redundant variables.
[0030] As an improvement to the present invention, step 4 is specifically as follows: Steps 2 and 3 convert the input multivariate time series into a directed graph G = (V, E) with node features. Next, a spectral-based graph neural network model is selected. Spectral-based GNNs perform convolution operations on the network in the Fourier domain by calculating the eigendecomposition of the Laplacian matrix, which represents the information propagation within the network. The convolution operation is the product between the input x and the filter g, where the filter g is a function of the Laplacian matrix.
[0031] y=g θ (L)x=g8(UΛU T )x=Ug g (Λ)xU T
[0032] Where U is a matrix whose column vectors are the unit eigenvectors of the network Laplacian matrix, and Λ is a diagonal matrix whose elements are the eigenvalues of the Laplacian matrix. The propagation rule of the graph convolutional neural network is
[0033]
[0034] Where l is the current layer number H l represents the nodes of layer l, W (l)is a learnable matrix parameter. The challenge faced by the spectrum-based graph neural network model is the asymmetry of the Laplace matrix. Inspired by the directed graph convolutional network model (DGCN) by constructing a symmetric Laplace matrix, the multivariate transfer entropy graph convolutional neural network (mTEGCN) is proposed. Based on the Perron-Frobenius theorem, the Laplace matrix of the graph is
[0035]
[0036] in A is the transition probability matrix of the directed graph, and Φ=diag(φ) is a diagonal matrix consisting of the Perron vector φ of P. In DGCN, the definition of the directed Laplacian matrix is
[0037] L sym =I-Φ 1 / 2 PΦ -1 / 2
[0038] The normalized Laplace operator is symmetric. Set up a multi-layer model for training. Taking two layers as an example, the information propagation method is expressed as
[0039]
[0040] Y=W Y (H)+B Y
[0041] in is the adjacency matrix of the graph; is the embedding representation of each node in the graph, calculated by the node embedding module; H is the output of the DGCN layer, W 1 and W 2 is the parameter matrix of each layer; Y is the final predicted value, calculated by linear transformation, W Y and B Y is a parameter. All parameters in directed graph convolutional networks are trained using the mean squared error (MSE) loss function. Spectral-based graph convolutional neural networks can avoid the high time complexity of traversing graph nodes and edge information, improving training efficiency.
[0042] Compared with the existing technology, the advantages of the present invention are as follows: the technical solution is based on a rutting prediction model based on multivariate transfer entropy and graph neural network. By calculating multivariate transfer entropy, the causal relationship between the variable and the rutting variable to be predicted can be better measured, and key variables can be identified, thereby avoiding the influence of redundant variables on the rutting variable to be predicted. The introduction of GRU and self-attention mechanism can better handle the influence of historical data of each variable on rutting data, thereby improving the generalization ability of the model. The selection of spectrum-based graph convolutional neural network can avoid the high computational complexity brought by traversing graph nodes and edges, thereby improving the computational efficiency of the model. Overall, compared with traditional statistics, machine learning methods, and deep learning methods, the accuracy and generalization ability of rutting prediction are improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a schematic diagram of the RIOHTRACK test section;
[0044] Figure 2 This is a framework diagram of the asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural network;
[0045] Figure 3 The following are the 19 experimental structures of asphalt pavement in RIOHTrack;
[0046] Figure 4 It is the overall flow chart of the present invention. DETAILED DESCRIPTION
[0047] In order to deepen the understanding of the present invention, this embodiment is described in detail below with reference to the accompanying drawings.
[0048] Example 1: See Figure 2 A method for predicting asphalt pavement rutting based on multivariate transfer entropy and graph neural network is proposed. The method comprises the following steps:
[0049] Step 1: Collect historical asphalt pavement service performance monitoring data and asphalt pavement performance influencing factors data, and classify according to the collected data;
[0050] Step 2: Convert the time series of rutting and multiple variables affecting rutting into a graph structure. The number of nodes in the graph is the number of variables, and the edges are defined as the causal relationship between the typical variables and the variables to be predicted, which is obtained by calculating the transfer entropy between the variables.
[0051] Step 3: Embed the nodes in the graph into a GRU and self-attention mechanism. This can be considered a pre-trained model.
[0052] Step 4: Use a spectrum-based graph neural network model to create a multi-layer model structure to train the variable node relationship graph obtained in the above steps and output a multi-step prediction sequence.
[0053] Step 1 is as follows: collect the monitoring data of asphalt pavement service performance and the data of factors affecting asphalt pavement performance over the years, classify the collected data, determine the input and output variables of the rutting prediction model, and use the time series of road rutting, deflection center value, deflection basin area, temperature, load and other data as the input of the model. The input time series I∈R n×t , where n is the number of rutting and other related variables, t is the size of each time series, and the output is multi-step prediction data about the road rutting data.
[0054] Step 2 is as follows: the variable n in step 1 is represented by a graph structure G = {V, E}, where V = {v1, v2, ...}, E = {e 1,2 ,e 1,3 ,…}, respectively represent the relationship between variables, where the edge relationship is obtained by calculating the multivariate transfer entropy. Transfer entropy is a non-parametric statistical method for measuring causal relationships between processes based on information theory and ordered Markov processes. Based on information entropy and calculated from conditional entropy, information entropy is used to measure the average level of uncertainty in a process. Given a process X t =x1,x2,…,x t , its information entropy is expressed as
[0055]
[0056] where p(x i ) is the variable component x i The probability of the process is greater, the greater the amount of information in the process, the greater the information entropy of the process. Conditional entropy is a measure of uncertainty in process Y given another process X. Consider two processes X t =x1,x2,…,x t and Y t =y1,y2,…,y t Represent the historical values of time series X and Y respectively. Then the conditional entropy of given X and Y can be expressed as
[0057]
[0058] Then the mutual information from process X to Y can be expressed as
[0059]
[0060] The reduction in the mean uncertainty of Y is quantified.
[0061] Xt =x τ-1 ,x τ-2 ,…,x τ-t and Y t =y τ-1 ,y τ-2 ,…,y τ-t As the characteristic process of X and Y, then from X t to Y t The transfer entropy can be expressed as
[0062]
[0063] in is the predicted value y τ and The conditional entropy of the joint distribution, It is y τ and historical sequence The conditional entropy of , k is the size of the sliding window. The transfer entropy is defined as τ ,VARIABLE X t and historical variable Y t The amount of information reduction, the value of transfer entropy has been shown to be equal to the value of conditional mutual information.
[0064] For a multivariate process X=X1,X2,…,X n The collective transfer entropy of Y is defined as the historical information of Y, each element subset Z (Z∈X n ) and the sum of other uncertainties:
[0065]
[0066] in It is an active information store used to capture historical information about Y; is a joint vector with a sliding window size of k; is the sum of the conditional mutual information increments, where each term is from the subset of accidental sources Z <n With information about X and Y, these subsets can be defined as U Y The inherent uncertainty of Y is due to the collective transfer entropy. Multivariate transfer entropy is introduced to capture causal relationships that manifest only through interactions, thereby reducing redundant links. Furthermore, a threshold can be set, setting transfer entropies below the threshold to zero. This reduces the influence of low-correlation variables and improves the method's generalization ability.
[0067] Step 3 is as follows: GRU is used to extract the temporal features of each input variable, which is regarded as the representation of the node embedding multivariate transfer entropy graph. The hidden unit of GRU is driven by LSTM unit, which captures the dependency relationship of different time series through the gate mechanism. Given an input time series X = x1, x2, ..., x t and the previous hidden state H = h1,h2,…,h t-1 , then the hidden state h t At time step t it is defined as
[0068] h t =Z t ⊙tanh(W h x t +U h (r t ⊙h t-1 )+(1-Z t Oh t-1 ))
[0069] z t =σ(W z x t +U z h t-1 ), r t =σ(W r x t +U r h t-1 )
[0070] where z t and r t are the update gate and reset gate respectively; σ and tanh are the logical sigmoid function and the tangent hyperbolic function; ⊙ refers to the multiplication of the corresponding elements of the two matrices; W h 、U h 、W z 、U z 、W r 、W r They are all learnable parameters. Update gate z t The output range is [0,1]. The closer it is to 1, the more information is remembered. Then use the attention-GRU based encoder module to generate the input time series represented as emb v ,v∈G is used as a node embedded in the multivariate transfer entropy graph. The output of the encoder at step n is
[0071]
[0072]
[0073] The vectors q, k, v are vector representations of the time steps in the intermediate matrices Q, K, V, which refer to the linear transformation of the input time series, i.e. Q = W Q X,K=W K X,V=W K X;α(q i ,k j ) refers to q i ,k j Similarity at time step i, j. The mean squared error (MSE) function is used for pre-training of this model.
[0074] Step 4 is as follows: Steps 2 and 3 transform the input multivariate time series into a directed graph G = (V, E) with node features. Next, a spectral-based graph neural network model is selected. Spectral-based GNNs perform convolution operations on the network in the Fourier domain by computing the eigendecomposition of the Laplacian matrix, which represents information propagation within the network. The convolution operation is the product of the input x and the filter g, which is a function of the Laplacian matrix.
[0075] y=g θ (L)x=g8(UΛU T )x=Ug g (Λ)xU T
[0076] Where U is a matrix whose column vectors are the unit eigenvectors of the network Laplacian matrix, and Λ is a diagonal matrix whose elements are the eigenvalues of the Laplacian matrix. The propagation rule of the graph convolutional neural network is
[0077]
[0078] Where l is the current layer number H l represents the nodes of layer l, W (l) is a learnable matrix parameter. The challenge faced by the spectrum-based graph neural network model is the asymmetry of the Laplace matrix. Inspired by the directed graph convolutional network model (DGCN) by constructing a symmetric Laplace matrix, the multivariate transfer entropy graph convolutional neural network (mTEGCN) is proposed. Based on the Perron-Frobenius theorem, the Laplace matrix of the graph is
[0079]
[0080] in A is the transition probability matrix of the directed graph, and Φ=diag(φ) is a diagonal matrix consisting of the Perron vector φ of P. In DGCN, the definition of the directed Laplacian matrix is
[0081] L sym=I-Φ 1 / 2 PΦ -1 / 2
[0082] The normalized Laplace operator is symmetric. Set up a multi-layer model for training. Taking two layers as an example, the information propagation method is expressed as
[0083]
[0084] Y=W Y (H)+B Y
[0085] in is the adjacency matrix of the graph; is the embedding representation of each node in the graph, calculated by the node embedding module; H is the output of the DGCN layer, W 1 and W 2 is the parameter matrix of each layer; Y is the final predicted value, calculated by linear transformation, W Y and B Y are parameters. All parameters in the directed graph convolutional network are trained by the mean squared error (MSE) loss function.
[0086] Specific embodiment: The data of this invention comes from China's first full-scale loop - RIOHTRACK. Figure 1 As shown, the loop is an elliptical closed curve, divided into two sections: a straight line section and a cylindrical section, with a total length of 2093 meters and a width of 15 meters. The main test section is set on the straight line and consists of 19 types of asphalt pavement structures, namely STR1-STR19. The pavement structure can be divided into 7 categories according to its structural combination. The thickness of the asphalt concrete structure layer in these pavements is 12, 18, 24, 28, 36, and 48 (52) m respectively. The detailed structures of the 19 types of asphalt pavements are shown in Figure 1. Figure 3 As shown. The track road is loaded by a closed trailer and runs for 12 hours every day during the day with an average speed of 40-50km / h. The track road performance test data includes real-time data and periodic data. Real-time data includes the collection of stress and strain information inside the structure, sensing of vehicle axle weight, and environmental monitoring inside and outside the structure. The vehicle axle weight and surface temperature are selected from the real-time data, and the center point deflection of the falling weight deflectometer (FWD) and the FWD deflection basin area are selected from the periodic data as features for rutting prediction. The vehicle axle load and pavement temperature are regarded as external factors affecting asphalt pavement rutting, and the FWD center point deflection and FWD deflection basin area that characterize the pavement material properties are regarded as internal factors affecting rutting. As shown in the following table
[0087] Table 1 Detailed information on rutting depth and influencing factors
[0088] variable name Variable meaning Numerical range unit Axle load Vehicle axle weight [4.45,7.62] lg(Ne) temperature Road surface temperature [-3.8,32.8] ℃ Depth of deflection basin Center point deflection of drop weight deflectometer [4.34,43.39] mm Deflection basin area Deflection basin area of the falling weight deflectometer [51.0,315.4] <![CDATA[mm 2 ]]> rutting Rutting depth [6.03,107.8] mm
[0089] First, the data of the above variables are preprocessed so that they are all in the same time series dimension, obtaining the input of our prediction method - the time series of 5 variables. The total number of available time series is 112, and the size of the lookback window is set to 12 time steps, sliding by 4 time steps, so the number of time windows is 25; then the transfer entropy between these 5 variables is calculated to obtain a 5×5 transfer entropy matrix, which is used as the weight information of the edge of the graph; secondly, GRU is used to extract the time features of the 5 variables, and the time series of the 5 variables is used as the input of GRU. The encoder module based on attention-GRU is used to generate the embedded information of the time series. The number of hidden layers and units is 2 and 16 respectively. The softmax function is used as the active function in the attention mechanism to select the information of the key variables, and the mean square error function (MSE) is used to adjust the parameters. Finally, the graph neural network in the mTEGCN proposed above is selected to train the graph structure obtained above. The number of GNN layers is set to 2, the number of convolution kernels in each layer is 1, and the number of hidden units is 64. The root mean square error (RMSE), mean absolute error (MAE) and determination coefficient R 2 To measure the predictive performance of the model, the first 80% of the data is used as the training and validation set, and the last 20% is used as the test set.
[0090] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting asphalt pavement rutting based on multivariate transfer entropy and graph neural network, characterized in that: The method comprises the following steps: Step 1: Collect the monitoring data of asphalt pavement service performance and the data of factors affecting asphalt pavement performance over the years, and classify them according to the collected data; Step 2: Convert the time series of rutting and multiple variables affecting rutting into a graph structure. The number of nodes in the graph is the number of variables. The edge is defined as the causal relationship between the typical variable and the variable to be predicted, which is obtained by calculating the transfer entropy between the variables. Step 3: Embed the nodes in the graph into a GRU and self-attention mechanism, which is a pre-trained model. Step 4: Use the spectrum-based graph neural network model to create a multi-layer model structure to train the variable node relationship graph obtained in the above steps and output a multi-step prediction sequence. Step 1 is as follows: collect the monitoring data of asphalt pavement service performance and the data of factors affecting asphalt pavement performance over the years, classify the collected data, determine the input and output variables of the rutting prediction model, and use the time series of pavement rutting, deflection center value, deflection basin area, temperature, and load data as the input of the model. The input time series I∈R n×t , where n is the number of rutting and other related variables, t is the size of each time series, and outputs multi-step prediction data about road rutting data; Step 2 is as follows: the variable n in step 1 is represented by a graph structure G = {V, E}, where V = {v1, v2, ...}, E = {e 1,2 ,e 1,3 ,…}, respectively represent the relationship between variables and variables, where the edge relationship is obtained by calculating the multivariate transfer entropy. Transfer entropy is a non-parametric statistical method for measuring the causal relationship between processes based on information theory and ordered Markov processes. It is based on information entropy and calculated from conditional entropy. Information entropy is used to measure the average level of uncertainty in the process. For a given process X t =x1,x2,…,x t , and its information entropy is expressed as Where p(x i ) is the variable component x i The probability of the process is greater, the greater the amount of information in the process, the greater the information entropy of the process. Conditional entropy is a measure of uncertainty in process Y given another process X. Consider two processes X t =x1,x2,…,x t and Y t =y1,y2,…,y t Represent the historical values of time series X and Y respectively, then the conditional entropy of given X and Y is expressed as Then the mutual information from process X to Y is expressed as The reduction in the average uncertainty of Y is quantified; X t =x τ-1 ,x τ-2 ,…,x τ-t and Y t =y τ-1 ,y τ-2 ,…,y τ-t As the characteristic process of X and Y, from X t To Y t The transfer entropy is expressed as in is the predicted value y τ and The conditional entropy of the joint distribution, Yes τ and historical sequence The conditional entropy of , k is the size of the sliding window, and the transfer entropy is defined as τ ,VARIABLE X t and the historical variable Y t The amount of information reduction, the value of transfer entropy has been proven to be equal to the value of conditional mutual information, For a multivariate process X = X1, X2, …, X n The collective transfer entropy of Y is defined as the incremental contribution of each element subset Z to the historical information of Y, Z∈X n , and the sum of other uncertainties: in It is an active information store used to capture Y's historical information; is a joint vector with a sliding window size of k; is the sum of the conditional mutual information increments, where each term comes from the accidental source subset Z <n to X and Y, these subsets are defined as U Y Based on the inherent uncertainty of collective transfer entropy, multivariate transfer entropy is introduced to capture the causal relationship that is only manifested through interaction, thereby reducing redundant links. By setting a threshold, the transfer entropy below the threshold is set to 0.
2. The asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural network according to claim 1 is characterized in that: Step 3 is as follows: GRU is used to extract the temporal features of each input variable, which is regarded as the representation of the node embedding multivariate transfer entropy graph. The hidden unit of GRU is driven by the long short-term memory (LSTM) unit, and the dependency of different time series is captured through the gate mechanism. Given an input time series X = x1, x2, ..., x t and the previous hidden state H = h1,h2,…,h t-1 , then the hidden state h t At time step t it is defined as z t =σ(W z x t +U z h t-1 ),r t =σ(W r x t +U r h t-1 ) where z t and r t are the update gate and reset gate respectively; σ and tanh are the logical sigmoid function and the tangent hyperbolic function; ⊙ refers to the multiplication of the corresponding elements of the two matrices; W h , U h , W z , U z , W r , U r They are all learnable parameters, and the update gate z t The output range is [0,1]. The closer it is to 1, the more information is remembered. Then, the attention-GRU based encoder module is used to generate the input time series represented as emb v ,v∈G is used as a node embedded in the multivariate transfer entropy graph, where the output of the encoder at step n is The vectors q, k, v are vector representations of the time steps in the intermediate matrices Q, K, V, which refer to the linear transformation of the input time series, i.e., Q = W Q X,K=W K X,V=W K X; α(q i ,k j ) refers to q i ,k j The similarity at time step i, j is used for pre-training of this model using the mean squared error function (MSE).
3. The asphalt pavement rutting prediction method based on multivariate transfer entropy and graph neural network according to claim 1 is characterized in that: Step 4 is as follows: Steps 2 and 3 convert the input multivariate time series into a directed graph G = (V, E) with node features, and select a spectral-based graph neural network model. The spectral-based GNN operates the convolution of the network in the Fourier domain by calculating the eigendecomposition of the Laplace matrix. The matrix represents the information propagation between networks. The convolution operation is the product between the input x and the filter g. The filter g is a function of the Laplace matrix. y=g θ (L)x=g θ (UΛU T )x=Ug θ (Λ)xU T Where U is a matrix whose column vector is the unit eigenvector of the network Laplacian matrix, Λ is a diagonal matrix whose elements are the eigenvalues of the Laplacian matrix, and the propagation rule of the graph convolutional neural network is Where l is the current layer number H l represents the nodes of layer l, W (l) is a learnable matrix parameter. Based on the Perron-Frobenius theorem, the Laplace matrix of the graph is in A is the transition probability matrix of the directed graph, Φ=diag(φ) is the diagonal matrix composed of the Perron vector φ of P. In DGCN, the definition of the directed Laplacian matrix is L sym =I-Φ 1 / 2 PF -1 / 2 The normalized Laplace operator is symmetric, and a multi-layer model is set for training. The information propagation method is expressed as Y=W Y (H)+B Y in is the adjacency matrix of the graph; is the embedding representation of each node in the graph, calculated by the node embedding module; H is the output of the DGCN layer, W 1 and W 2 is the parameter matrix of each layer; Y is the final prediction value, calculated by linear transformation, W Y and B Y are parameters, and all parameters in directed graph convolutional networks are trained by the mean squared error (MSE) loss function.
