Neural tensor ring fusion method for traffic flow data completion

Abstract

The invention discloses a neural tensor ring fusion method for traffic flow data completion, which belongs to the field of traffic data completion, and comprises the following steps: step 1, acquiring traffic flow data in a road network system, and modeling the traffic flow data into a third-order tensor; 2, constructing an objective function of a neural tensor ring fusion model by referring to a tensor ring decomposition thought; and step 3, traffic flow data completion is carried out based on the neural tensor ring fusion model. According to the method, a neural tensor ring fusion model is constructed by adopting a tensor ring decomposition thought, potential features of data are learned by adopting a three-order factor tensor form, and the structure and information of original data can be better stored while high-order data tensor is expressed and processed; meanwhile, multi-dimensional interaction of traffic data in time is captured by adopting a convolutional long-short-term memory network, and missing data are complemented together by using a plurality of matrixes containing historical time characteristics.

Description

Technical Field

[0001] The present invention belongs to the field of traffic data completion, and in particular relates to a neural tensor loop fusion method for traffic flow data completion. Background Art

[0002] With the continuous development of information collection technology, the number of IoT sensors deployed by intelligent transportation systems (ITS) in cities has gradually increased. Sensors can collect traffic flow data at all times, which can not only provide urban residents with richer road network information, but also help ITS solve problems such as urban pollution and traffic congestion caused by large-scale urban growth. However, due to many reasons such as weather conditions, hardware failures, power outages, etc., traffic data is often missing. The conclusions and decisions drawn by researchers using these unprocessed missing traffic data may be significantly biased. Therefore, how to accurately complete traffic flow data has become a hot issue of concern to researchers.

[0003] The process of traffic data completion is essentially to capture the features of known data and their spatiotemporal correlations through models, and use the learned data features to fill in the missing data. Therefore, whether the hidden features and their relationships in the data can be fully captured is the key to the traffic completion model. As a high-order extension of the matrix decomposition method, researchers have tried to use tensor decomposition models to solve the problem of traffic data completion. This type of model can effectively extract the linear features of the data and store the features in a small-scale factor matrix (tensor), which improves the completion accuracy while also being interpretable. However, the tensor completion method focuses on the learning of linear features, and its ability to complete data with high-order and complex nonlinear correlations is very limited.

[0004] In order to improve the above shortcomings, some researchers have extended the tensor decomposition model to the form of neural networks, that is, to construct tensor neural networks (TNNs). TNNs not only retain the advantages of the tensor decomposition model, but also have strong expression and processing capabilities for high-dimensional data, and have excellent performance in the field of data completion. Despite this, the neural tensor network model currently used for traffic flow data completion still has the following problems: First, the existing neural tensor network model refers to the idea of ​​tensor CP decomposition, and mostly uses the second-order factor matrix form to learn the potential features in the data, which makes the model unable to capture the hidden high-order nonlinear relationship between features, and the completion accuracy needs to be further improved; second, the existing neural tensor network model only uses a single vector (matrix) containing the current time feature to complete the missing data, ignoring the impact of historical time features on the data, and cannot well represent the dynamic law of traffic flow changes over time. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a neural tensor ring fusion method for traffic flow data completion. Aiming at the problem that the existing methods cannot fully learn the high-order nonlinear relationship between data features, the present invention adopts the tensor ring (TR) decomposition idea to construct a new neural tensor network, namely the neural tensor ring fusion model. The model learns the potential features of the data in the form of a third-order factor tensor, and can better preserve the structure and information of the original data while expressing and processing high-order data tensors. At the same time, in view of the problem that the existing model cannot well characterize the dynamic law of traffic flow changing over time, the present invention adopts a convolutional long short-term memory network (ConvLSTM) to capture the multi-dimensional interaction of traffic data in time, and uses multiple matrices containing historical time features to jointly complete the missing data.

[0006] The technical solution of the present invention is as follows: A neural tensor loop fusion method for traffic flow data completion includes the following steps: Step 1: Obtain traffic flow data in the road network system and model the traffic flow data as a third-order tensor; Step 2: Construct the objective function of the neural tensor ring fusion model based on the tensor ring decomposition idea; Step 3: Complete traffic flow data based on the neural tensor loop fusion model.

[0007] Furthermore, in step 1, the traffic flow data is modeled as a third-order tensor ,in Represents the number of nodes in the road network system, Indicates the time interval, is the total number of days in the data set; the third-order tensor consists of several elements, one element represents a traffic flow value, and the definition element Indicates Tianzhongdi Period Traffic flow value at each node.

[0008] Furthermore, in step 2, the tensor ring decomposition idea refers to decomposing a high-order tensor into a series of linear multiplications of third-order factor tensors, which are connected end to end to form a ring structure; A high-order tensor of , tensor ring decomposition can transform The high-order tensor of order is decomposed into A third-order factor tensor; For high-order tensors The dimension of the order; the tensor ring decomposition formula is: ; in, is a high-order tensor The median coordinate is Elements of For high-order tensors The coordinates of the elements on the order; represents the trace operation; is the order number; is composed of high-order tensors The coordinates of the elements on the order The determined matrix is ​​taken from Factor Tensor On the second level slices. In the tensor, the coordinates of the elements on the order determine which slice to take.

[0009] Furthermore, in step 2, the objective function of the neural tensor loop fusion model is expressed as: ; in, , , are three different third-order factor tensors, , , are different artificial hyperparameters, representing the tensor ring ranks of the first factor tensor, the second factor tensor, and the third factor tensor, respectively. , , They are the dimensions of the 1st, 2nd, and 3rd order of the high-order tensor respectively; Represents all parameters that need to be trained in the network; is the element coordinate index; After completion Tianzhongdi Period Traffic flow value at each node; , , are different regularization coefficients.

[0010] Furthermore, in step 3, the neural tensor ring fusion model includes an embedding layer and an interaction layer; the embedding layer constructs three third-order factor tensors to fully learn the potential features in the original traffic data, and uses the ConvLSTM encoder to characterize the temporal interaction of traffic data; in the interaction layer, the model converts each matrix containing different features into a fused feature through vectorization and splicing operations, and then gradually mines the complex nonlinear correlation between features through the feature extraction layer.

[0011] Furthermore, the specific process of step 3 is as follows: Step 3.1: embed data for each traffic flow value; the specific process is as follows: Each traffic flow value contains three items: node, time period, and day. These three items are encoded one-hot to construct the one-hot vector of node, time period, and day: ; ; ; in, For the One-hot vector of nodes; For the One-hot vector for each time period; For the One-hot vector of day; represents the one-hot encoding process; is the time scale parameter; Through the one-hot vector, the model obtains embedding matrices of different orders. Each embedding matrix contains its own potential features. The embedding matrix of the node contains the node features, the embedding matrix of the time period contains the time features, and the embedding matrix of the day contains the day features. Each embedding matrix is ​​as follows: ; ; ; in, For the The embedding matrix of nodes; For the The embedding matrix of time periods; , Respectively Day, The embedding matrix of the day; Represents the mode-2 product of a tensor; Before The embedding matrix is ​​input into the ConvLSTM encoder to obtain the Potential characteristics of the day: ; ; in, , , , They are initial, first, and , memory cells; , , , They are initial, first, and , hidden states; is the ConvLSTM encoder; is the first The underlying characteristics of the sky; is the Sigmoid activation function; , Respectively represent the weight matrix and partial rank term of the mapping layer; Step 3.2: Use the embedding matrix to construct fusion features. Specifically, the node features, time period features, and day features in the embedding matrix are vectorized and concatenated to obtain fusion features. The formula is as follows: ; ; ; ; in, , , They are vectorized node features, time period features, and day features respectively; Represents vectorization; Represents a splicing operation; To fuse features, all feature elements are included. is the set hyperparameter, representing the potential feature dimension; Step 3.3: Model adoption The feature extraction modules capture the nonlinear correlation between features. Each feature extraction module consists of a fully connected layer and a one-dimensional convolutional layer. The process is expressed as: ; in, represents the feature extraction module; , , Respectively represent the first and , The output vector of the feature extraction module; Step 3.4: The neural tensor loop fusion model will Input the fully connected layer to get the final completed Tianzhongdi Period Traffic flow value at each node : ; in, represents a fully connected layer; , are the weight matrix and partial rank term of the fully connected layer respectively.

[0012] The beneficial technical effects brought by the present invention: In view of the problems existing in the current traffic data completion neural tensor network model, the present invention proposes a new neural tensor ring fusion framework based on the TR decomposition method; this method uses the third-order factor tensor in the embedding layer to learn the different features implicit in the data, and fuses the different features in the feature interaction layer, and then uses the feature extraction module to capture the correlation of the fusion vector, so as to fully explore the hidden high-order nonlinear features between the features. In addition, through the ConvLSTM encoder, the model can more comprehensively learn the potential dynamic time characteristics of the data, thereby effectively utilizing the dynamic correlation of the data in time. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The figure is a flow chart of the neural tensor loop fusion method for traffic flow data completion of the present invention.

[0014] Figure 2 It is a line graph of the root mean square error of each model in the random missing scenario in the experiment of the present invention.

[0015] Figure 3 It is a line graph of the root mean square error of each model in the non-random missing scenario in the experiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: The present invention believes that the traffic flow tensor has correlation in the three modes of "node", "time period" and "day". Based on the above viewpoints, the neural tensor ring fusion framework proposed in the present invention refers to the idea of ​​third-order tensor ring decomposition, uses three factor tensors in the embedding layer to learn different features of the data respectively, and fuses the features through the interaction layer, so as to effectively capture and utilize the high-order nonlinear relationship between data features and achieve high-precision data completion.

[0017] The present invention considers that the change of traffic flow over time is smooth. Based on this view, the present invention proposes a new method for constructing a time feature matrix, which inputs multiple matrices containing historical time features into ConvLSTM to generate a time feature matrix of corresponding elements, effectively utilizing the dynamic correlation of data in time.

[0018] like Figure 1 As shown, the present invention comprises the following steps: Step 1: Obtain traffic flow data in the road network system and model the traffic flow data as a third-order tensor ,in Represents the number of nodes in the road network system, Indicates the time interval, is the total number of days in the data set; the third-order tensor consists of several elements, one element represents a traffic flow value, and the definition element Indicates Tianzhongdi Period Traffic flow value at each node.

[0019] Due to equipment failure or communication problems, traffic flow data often has a certain degree of missing. The goal of this invention is to build a complementary model in the absence of data to effectively infer Missing traffic flow values ​​in .

[0020] Step 2: Construct the objective function of the neural tensor ring fusion model based on the tensor ring decomposition idea; Tensor ring decomposition can effectively reduce the dimension and represent high-order tensors. It can decompose a high-order tensor into a series of linear multiplications of third-order factor tensors. These tensors are connected end to end to form a ring structure. For example, for A high-order tensor of , tensor ring decomposition can transform The high-order tensor of order is decomposed into A third-order factor tensor, The factor tensor is named , In the above expression, and are artificially specified hyperparameters, respectively referred to as The tensor ring rank and the The tensor ring rank of a factor tensor directly determines The number of latent features in ; For high-order tensors The dimension of the order, For high-order tensors The dimension of the high-order tensor The coordinates of the elements on the order are , and convert the high-order tensor The median coordinate is The elements of , For high-order tensors The element coordinates of the order can be decomposed into the following formula through the tensor ring: ; in, represents the trace operation; is composed of high-order tensors The coordinates of the elements on the order The determined matrix is ​​taken from Factor Tensor On the second level In a tensor, the coordinates of the elements on the order determine which slice to take.

[0021] The neural tensor loop fusion model framework for traffic data completion constructed by the present invention is as follows: Figure 1 As shown in the figure, the model builds a multi-layer neural network architecture with reference to the above tensor ring decomposition idea, which is mainly composed of two main layers: the embedding layer and the interaction layer. In the embedding layer, the model constructs three third-order factor tensors to fully learn the potential features in the original traffic data, and uses the ConvLSTM encoder to represent the temporal interaction of traffic data; in the interaction layer, the model transforms each matrix containing different features into fused features through vectorization and splicing operations, and then gradually mines the complex nonlinear correlations between features through the feature extraction layer.

[0022] The objective function of the neural tensor loop fusion model can be expressed as: ; in, , , are three different third-order factor tensors, , , are different artificial hyperparameters, representing the tensor ring ranks of the first factor tensor, the second factor tensor, and the third factor tensor, respectively. , , are the dimensions of the 1st, 2nd, and 3rd order of the high-order tensor, respectively. For the sake of introduction, the three are equal here, and the sizes of the three are set to the potential feature dimensions, that is, , is the set hyperparameter, representing the potential feature dimension; Represents all parameters that need to be trained in the network; is the element coordinate index; After completion Tianzhongdi Period Traffic flow value at each node; , , For different regularization coefficients, add regularization terms here To prevent the model from overfitting.

[0023] Step 3: Complete the traffic flow data based on the neural tensor loop fusion model; the specific process is as follows: Step 3.1, embed data for each traffic flow value; Referring to the idea of ​​tensor ring decomposition, three third-order factor tensors are constructed in the embedding layer, so that the model can fully learn the potential spatiotemporal characteristics of traffic flow data at all orders during training. In addition, the model of the present invention uses the ConvLSTM encoder to enhance the model's ability to capture dynamic time correlations. The specific process is: Each traffic flow value contains three items: node, time period, and day. The IDs of these three items are uniquely encoded (the unique encoding is a string consisting of 0 and 1) to construct unique vectors of nodes, time periods, and days: ; ; ; in, For the One-hot vector of nodes; For the One-hot vector for each time period; For the One-hot vector of day; represents the one-hot encoding process; here is an artificially specified hyperparameter, called the time scale parameter, which determines the number of units of the ConvLSTM encoder below, that is, The characteristics of the day are past The characteristics of the sky are jointly determined.

[0024] Through the one-hot vector, the model can obtain embedding matrices at different orders. Each embedding matrix contains its own potential features. The embedding matrix of the node contains the node features, the embedding matrix of the time period contains the time features, and the embedding matrix of the day contains the day features. Each embedding matrix is ​​as follows: ; ; ; in, For the The embedding matrix of nodes; For the The embedding matrix of time periods; , Respectively Day, The embedding matrix of the day; Represents the mode-2 product of tensors.

[0025] It is worth noting that in order to further extract the time pattern of traffic flow data changing with the number of days, we give up directly using To indicate the The potential characteristics of the day, but the previous The embedding matrix is ​​input into the ConvLSTM encoder to obtain a more accurate Potential characteristics of the day: ; ; in, , , , They are initial, first, and , memory cells; , , , They are initial, first, and , hidden states; is the ConvLSTM encoder; is the first The underlying characteristics of the sky; is the Sigmoid activation function; , They represent the weight matrix and partial rank term of the mapping layer respectively.

[0026] Step 3.2: Use the embedding matrix to construct fusion features. Specifically, the node features, time period features, and day features in the embedding matrix are vectorized and concatenated to obtain fusion features. The formula is as follows: ; ; ; ; in, , , They are vectorized node features, time period features, and day features respectively; Represents vectorization; Represents a splicing operation; To fuse features, all feature elements are included.

[0027] Step 3.3: The model of the present invention adopts The feature extraction modules capture the nonlinear correlation between features. Each feature extraction module consists of a fully connected layer and a one-dimensional convolutional layer. The process is expressed as: ; in, represents the feature extraction module; , , Respectively represent the first and , The output vector of the feature extraction module.

[0028] Step 3.4: Finally, the neural tensor loop fusion model will Input the fully connected layer to get the final completed Tianzhongdi Period Traffic flow value at each node : ; in, represents a fully connected layer; , are the weight matrix and partial rank term of the fully connected layer respectively.

[0029] In summary, the interaction layer function adopted by the present invention can be expressed as: ; in, Indicates continuous use A feature extraction module; Thanks to this construction, the interaction layer function of the neural tensor ring fusion model can learn as the input data changes, fully tapping the hidden features of the data and the high-order complex correlations between the features, thereby achieving better experimental accuracy.

[0030] In order to prove the feasibility and superiority of the present invention, the following comparative experiments are given. The data set selects the PEMS04 highway traffic data set, including two missing scenarios: random missing and non-random missing, with missing rates of 20%, 40%, 60% and 80%. The comparison models select TR (tensor ring decomposition model), NTF (neural tensor network model based on tensor CP decomposition idea) and NTC (neural tensor network model using three-dimensional convolutional neural network to extract features), and the evaluation index selects the root mean square error. The comparison results are shown in Figure 2. Figure 2 , Figure 3 shown. Figure 2 , Figure 3The line graphs are respectively the root mean square errors of the models in the random missing and non-random missing scenarios in the experiment of the present invention. It can be clearly seen that the root mean square error of the present invention achieves the minimum value in all scenarios and all missing rates, indicating that the present invention can achieve higher prediction accuracy.

[0031] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A neural tensor loop fusion method for traffic flow data completion, characterized in that: The steps include: Step 1: Obtain traffic flow data in the road network system and model the traffic flow data as a third-order tensor; Step 2: Construct the objective function of the neural tensor ring fusion model based on the tensor ring decomposition idea; Step 3: Complete traffic flow data based on the neural tensor loop fusion model.

2. The neural tensor loop fusion method for traffic flow data completion according to claim 1, characterized in that: In step 1, the traffic flow data is modeled as a third-order tensor ,in Represents the number of nodes in the road network system, Indicates the time interval, is the total number of days in the data set; the third-order tensor consists of several elements, one element represents a traffic flow value, and the definition element Indicates Tianzhongdi Period Traffic flow value at each node.

3. The neural tensor loop fusion method for traffic flow data completion according to claim 2 is characterized in that: In step 2, the idea of ​​tensor ring decomposition is to decompose a high-order tensor into a series of linear multiplications of third-order factor tensors, which are connected end to end to form a ring structure; A high-order tensor of , tensor ring decomposition can transform The high-order tensor of order is decomposed into A third-order factor tensor; For high-order tensors The dimension of the order; the tensor ring decomposition formula is: ； in, is a high-order tensor The median coordinate is Elements of For high-order tensors The coordinates of the elements on the order; represents the trace operation; is the order number; is composed of high-order tensors The coordinates of the elements on the order The determined matrix is ​​taken from Factor Tensor On the second level slices. In the tensor, the coordinates of the elements on the order determine which slice to take.

4. The neural tensor loop fusion method for traffic flow data completion according to claim 3 is characterized in that: In step 2, the objective function of the neural tensor loop fusion model is expressed as: ； in, , , are three different third-order factor tensors, , , are different artificial hyperparameters, representing the tensor ring ranks of the first factor tensor, the second factor tensor, and the third factor tensor, respectively. , , They are the dimensions of the 1st, 2nd, and 3rd order of the high-order tensor respectively; Represents all parameters that need to be trained in the network; is the element coordinate index; After completion Tianzhongdi Period Traffic flow value at each node; , , are different regularization coefficients.

5. The neural tensor loop fusion method for traffic flow data completion according to claim 4 is characterized in that: In step 3, the neural tensor ring fusion model includes an embedding layer and an interaction layer; the embedding layer constructs three third-order factor tensors to fully learn the potential features in the original traffic data, and uses the ConvLSTM encoder to characterize the temporal interaction of traffic data; in the interaction layer, the model converts each matrix containing different features into a fused feature through vectorization and concatenation operations, and then gradually mines the complex nonlinear correlation between features through the feature extraction layer.

6. The neural tensor loop fusion method for traffic flow data completion according to claim 5, characterized in that: The specific process of step 3 is as follows: Step 3.1: embed data for each traffic flow value; the specific process is as follows: Each traffic flow value contains three items: node, time period, and day. These three items are encoded one-hot to construct the one-hot vector of node, time period, and day: ； ； ； in, For the One-hot vector of nodes; For the One-hot vector for each time period; For the One-hot vector of day; represents the one-hot encoding process; is the time scale parameter; Through the one-hot vector, the model obtains embedding matrices of different orders. Each embedding matrix contains its own potential features. The embedding matrix of the node contains the node features, the embedding matrix of the time period contains the time features, and the embedding matrix of the day contains the day features. Each embedding matrix is ​​as follows: ； ； ； in, For the The embedding matrix of nodes; For the The embedding matrix of time periods; , Respectively Day, The embedding matrix of the day; Represents the mode-2 product of a tensor; Before The embedding matrix is ​​input into the ConvLSTM encoder to obtain the Potential characteristics of the day: ； ； in, , , , They are initial, first, and , memory cells; , , , They are initial, first, and , hidden states; is the ConvLSTM encoder; is the first The underlying characteristics of the sky; is the Sigmoid activation function; , Respectively represent the weight matrix and partial rank term of the mapping layer; Step 3.2: Use the embedding matrix to construct fusion features. Specifically, the node features, time period features, and day features in the embedding matrix are vectorized and concatenated to obtain fusion features. The formula is as follows: ； ； ； ； in, , , They are vectorized node features, time period features, and day features respectively; Represents vectorization; Represents a splicing operation; To fuse features, all feature elements are included. is the set hyperparameter, representing the potential feature dimension; Step 3.3: Model adoption The feature extraction modules capture the nonlinear correlation between features. Each feature extraction module consists of a fully connected layer and a one-dimensional convolutional layer. The process is expressed as: ； in, represents the feature extraction module; , , Respectively represent the first and , The output vector of the feature extraction module; Step 3.4: The neural tensor loop fusion model will Input the fully connected layer to get the final completed Tianzhongdi Period Traffic flow value at each node : ； in, represents a fully connected layer; , are the weight matrix and partial rank term of the fully connected layer respectively.

Images