Method and System for Classifying Truck Travel Categories Based on Multi-Scale Hypergraph Convolution Deep Clustering

Abstract

The present invention relates to the field of transportation information engineering, and discloses a method and system for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering. The method includes: constructing a truck travel feature matrix at each spatial scale; constructing a truck travel pattern hypergraph at each spatial scale to model the "many-to-many" high-order correlation relationship of trucks at multiple spatial scales; constructing a multi-scale hypergraph convolutional deep clustering model for training and testing the classification of truck travel categories; and conducting an example verification of the classification of truck travel categories. Applying the proposed model to a series of empirical datasets and self-collected datasets proves that it can achieve good classification performance superior to the baseline model.

Description

Technical Field

[0001] The present invention relates to the technical field of transportation information engineering, and particularly relates to a method and system for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering. Background Art

[0002] The accurate classification of group categories is an important basis for deeply understanding group differences, formulating policy opinions by classification, and initiating emergency responses precisely. When completing freight transportation operations, trucks continuously travel across multi-scale spaces such as bus stations, streets, and administrative regions, and thus there are "many-to-many" associations and cross-scale interactive travel behaviors at each scale. Therefore, when classifying truck travel mode categories, it is necessary to fully consider the "many-to-many" high-order association characteristics of travel modes among trucks with multi-space scale interactions.

[0003] Existing research has used classical algorithms such as k-means, DBSCAN, and hierarchical clustering to classify truck travel mode categories, but has not utilized the personalized travel association characteristics among trucks to affect the clustering results. In addition, graph clustering algorithms have also been used in related technologies to learn the implicit feature representations behind the original data, capture the implicit association structures among individuals, and provide a breakthrough for group category classification. However, the traditional graph's way of "connecting two points with a single edge" is not sufficient to express the "many-to-many" mapping attributes between vertices, which hinders the performance of group category classification to a certain extent, and thus will lead to inaccurate classification of truck group categories. Summary of the Invention

[0004] The present invention provides a method and system for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering to solve the defect that the group classification of trucks in related technologies is inaccurate. In the solution of this application, the travel records of trucks can be clustered by constructing a hypergraph, which improves the accuracy of group category classification.

[0005] The present invention provides a method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering, including:

[0006] Step 1, construct the truck travel feature matrices at each spatial scale. According to the historical travel record set of N trucks, obtain the historical travel record set of N trucks at a given time period; at spatial scale s, extract F-dimensional travel features for each truck, and construct the travel feature matrix of N trucks Obtain the truck travel feature matrices X = {X1, …, X s , …, X S} at S hierarchical spatial scales, where x s,i is the travel feature expression of the i-th truck at spatial scale s, 1 ≤ i ≤ N; x iis the travel feature expression of the \(i\)-th truck, where \(1\leq i\leq N\); the \(F\)-dimensional travel features of each truck come from the time, space, and attribute dimensions. The time dimension includes the number of times the truck driver drives while fatigued, the duration of night driving, the departure time of the truck, and the average daily travel duration; the space dimension includes the travel distance of the truck, the function of the truck stop point, and the average travel speed of the truck; the attribute dimension includes the license plate color and the number of transportation trips.

[0007] Step 2: Construct the truck travel pattern hypergraph at each spatial scale, model the many-to-many high-order association relationship of trucks, and according to the travel feature matrix \(X\) of the trucks in Step 1 s , construct \(X\) s The travel pattern hypergraph \(G\) at the spatial scale \(s\) s =(V s , \(X\) s , \(W\) s , \(E\) s , \(H\) s ), regard \(N\) trucks as \(N\) nodes on the graph, and form the node set \(V\) representing all trucks s ; regard the \(F\)-dimensional travel features of each truck as the attributes of each node, and form the node attribute set \(X\) representing the truck travel pattern s ; connect two adjacent nodes, and use the reciprocal of their Euclidean distance as the edge weight value corresponding to the nodes, and form the edge weight value set \(W\) representing the importance degree of the neighborhood group s ; select the top \(k - 1\) neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the \(i\)-th truck Analogize the above neighborhood group to a hyperedge, obtain a set of \(M\) hyperedges, and form the hyperedge set \(E\) representing the complex association relationship of the many-to-many travel patterns of the neighborhood group s ; \(H\) s is an \(N\times N\) association matrix representing the membership relationship between a single truck and the neighborhood group. The internal element \(h(v, e)\) of \(H\) s characterizes the membership relationship between node \(v\) and hyperedge \(e\). The specific calculation method is shown in Equation (1):

[0008]

[0009] where \(v\in e\) means the node is on the hyperedge;

[0010] Step 3: Construct a multi-scale hypergraph convolutional deep clustering model for training and testing the classification of truck travel categories, including:

[0011] Step 3.1, use the autoencoder model to learn the implicit travel pattern representation of the trucks;

[0012] Step 3.2, use the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck neighborhood group in the local space;

[0013] Step 3.3: Using the cross - self - attention mechanism, fuse the implicit travel pattern representations of the truck and its neighborhood groups at each spatial scale, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale;

[0014] Step 3.4: Train and test the performance of truck travel category classification;

[0015] Step 4: Instance verification. Input the travel feature matrix X of N trucks at the (S + 1)-th spatial scale S+1 , construct the hypergraph G′ of truck travel patterns that fuses multi - scale spatial features S+1 , and achieve the classification of truck group travel categories; Apply it to the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective classification performance of the truck travel category classification method based on multi - scale hypergraph convolutional deep clustering through comparison and ablation experiments.

[0016] According to the truck travel category classification method based on multi - scale hypergraph convolutional deep clustering provided by the present invention, the use of the auto - encoder model to learn the implicit travel pattern representation of the truck includes:

[0017] Introduce the auto - encoder model. Through the built - in encoder, learn the implicit travel pattern representation of the truck travel feature matrix X at the spatial scale s in Step 1 s at the convolutional layer l (1 ≤ l ≤ L) which is calculated by formula (2):

[0018]

[0019] where σ ReLU is the linear activation function; AutoEncoder is the auto - encoder. When l = 1, is the truck travel feature matrix X output in Step 1 s ; W (l) is the convolutional kernel weight of the l - th layer, and b (l) is the bias value of the l - th layer. Both are weight parameters to be trained in the model;

[0020] Through the decoder built in the auto - encoder, after L convolutional layers, obtain the reconstructed to complete the decoding operation; Minimize the average reconstruction error generated by the auto - encoder model at all spatial scales expressed by formula (3) to obtain the truck travel pattern representation closest to the real situation:

[0021]

[0022] where, ||·|| Fro is the F - norm;

[0023] For the spatial scale s, according to the truck travel characteristic matrix X s 's distribution, obtain the probability distribution curve GT of truck travel categories in the real scenario; according to the hidden feature representation of the truck travel pattern in the L-th convolutional layer 's distribution, obtain the probability distribution curve TR of truck travel categories obtained by the autoencoder partition, and measure the cumulative error between curve GT and curve TR at all spatial scales Calculated by formula (4):

[0024]

[0025] where GT sij and AE sij are respectively the probabilities that truck i belongs to category j in curve GT and curve AE at spatial scale s.

[0026] According to the method for dividing truck travel categories based on multi-scale hypergraph convolutional deep clustering provided by the present invention, the use of the hypergraph convolutional network structure to learn the hidden travel pattern representation of the truck neighborhood group in the local space includes:[[]]

[0027] According to the truck travel pattern hypergraph G at any spatial scale s obtained in step 2 s , perform hypergraph convolutional network operations to capture the local spatial topological changes of the truck neighborhood group. For the l-th convolutional layer, the hidden travel pattern of the truck neighborhood group Calculated by formula (5):

[0028]

[0029] where Θ is the convolutional hyperparameter; when l = 1, is the truck travel characteristic matrix X obtained in step 1 s ; H s Tr is the transpose matrix of H s ; D s,v = [d(v)] is the degree matrix of node v at spatial scale s, used to describe the importance of node v in the hypergraph G s ; the internal element d(v) is the sum of the weights of all hyperedges connected to this vertex v, calculated by formula (6):

[0030]

[0031] D s,e = [d(e)] is the degree matrix of hyperedge e at spatial scale s, used to describe the importance of hyperedge e in the hypergraph G sThe importance level in; the internal element d(e) is the sum of the number of all nodes connected by the hyperedge e, which is calculated by formula (7):

[0032]

[0033] According to the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by the present invention, the cross self-attention mechanism is used to fuse the implicit travel pattern representations of the aforementioned trucks and their neighborhood groups at each spatial scale, and deep clustering is performed through a hypergraph convolutional network to obtain the division results of truck travel categories at each spatial scale, including:

[0034] According to the truck implicit travel pattern representation in the l-th convolutional layer in step 3.1 In step 3.2, the implicit travel pattern representation of the truck neighborhood group in the l-th convolutional layer Introduce the cross self-attention mechanism, and obtain the implicit travel pattern expression of the truck neighborhood group in the l-th convolutional layer at the s-th spatial scale through formula (8) Then, obtain a new expression that fuses the travel patterns of the truck neighborhood group in the l-th convolutional layer at all spatial scales through formula (9)

[0035]

[0036] In formula (8), CrossSelfAttention represents the cross self-attention mechanism; is the n-th vector sequence of; α mn is the m-th query and the n-th key The attention weight, indicating their correlation β mn is the m-th query and the n-th key The attention weight, indicating their correlation and is the attention scoring function, which is used to measure the correlation between the two in the formula and is quantitatively expressed using the dot product model;

[0037] Based on the fusion features of multiple spatial scales, update the corresponding hypergraphs at each spatial scale Correspondingly splice to obtain a dynamic hypergraph As the input of the hypergraph convolutional network in the (l + 1)-th convolutional layer and the output result corresponding to the autoencoder in the (l + 1)-th layer After performing cross - self - attention fusion, perform hypergraph convolutional network operations, and obtain the probability distribution result of truck travel categories in the (l + 1)-th convolutional layer through the following formula (10). Finally, complete the dynamic division of truck travel categories;

[0038]

[0039] In formula (10), W′ s 、H′ s 、D′ s,v 、D′ s,e are respectively the hyper - edge weight set, incidence matrix, node degree matrix, and hyper - edge degree matrix of multiple spatial - scale fusion hypergraphs ;

[0040] For any spatial scale s, according to the distribution of the truck implicit travel pattern representation in the L - th convolutional layer, obtain the probability distribution curve HC of truck travel categories obtained by hypergraph clustering at spatial scale s, and measure the cumulative error between curve HC and curve GT at all spatial scales Calculated by formula (11):

[0041]

[0042] where HC sij refers to the probability that truck i belongs to category j in curve HC at spatial scale s.

[0043] According to the truck travel category division method based on multi - scale hypergraph convolutional deep clustering provided by the present invention, the training and testing of the performance of truck travel category division include:

[0044] From formula (12), based on the average reconstruction error and the and the obtain the overall error

[0045]

[0046] where η1, η2, and η3 are the weight coefficients of the three - item errors respectively;

[0047] Continuously loop and iterate the above steps 3.1 - 3.3 to minimize the cumulative error until the value no longer changes, the model converges, and at this time, the divided truck travel categories are truly available, completing the training process of the model;

[0048] Apply the trained model to the test dataset to verify the performance of the model in dividing truck travel categories for any spatial scale.

[0049] According to the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by the present invention, through comparative and ablation experiments, the effective division performance of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering is verified, including:

[0050] The effective division performance of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering is verified through four indicators: accuracy, normalized virtual information, average Rand index, and F1-score.

[0051] The present invention also provides a truck travel category division system based on multi-scale hypergraph convolutional deep clustering, including:

[0052] A truck travel feature matrix construction module, which is used to obtain the historical travel record set of N trucks at a given time period according to the historical travel record set of N trucks; at the spatial scale s, extract F-dimensional travel features for each truck and construct the travel feature matrix X of N trucks s ;

[0053] A truck travel mode hypergraph construction module, which is used to construct the travel mode hypergraph G of X at the spatial scale s according to the travel feature matrix X of the truck s , construct X s The travel mode hypergraph G at the spatial scale s = (V s , X s , W s , E s , H s ), take N trucks as N nodes on the graph, and form a node set V representing all trucks s ; regard the F-dimensional travel features of each truck as the attributes of each node, and form a node attribute set X representing the truck travel mode s ; connect two adjacent nodes, and use the reciprocal of the Euclidean distance between the two as the edge weight value of the corresponding nodes, and form an edge weight value set W representing the importance degree of the neighborhood group s ; select the first k - 1 neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the i-th truck s ; analogize the above neighborhood group to a hyperedge, obtain a set of M hyperedges, and form a hyperedge set E representing the complex association relationship of the many-to-many travel mode of the neighborhood group , H s is an N × N association matrix representing the membership relationship between a single truck and the neighborhood group; s

[0054] An implicit travel mode representation learning module for trucks, which is used to utilize the hypergraph convolutional network structure to learn the implicit travel mode representation of the truck neighborhood group in the local space;

[0055] The neighborhood group implicit travel pattern representation learning module is used to utilize the cross self-attention mechanism to fuse the implicit travel pattern representations of trucks and their neighborhood groups at various spatial scales, and perform deep clustering through a hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale;

[0056] The hypergraph convolutional deep clustering module is used to utilize the cross self-attention mechanism to fuse the implicit travel pattern representations of trucks and their neighborhood groups at various spatial scales, and perform deep clustering through a hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale;

[0057] The training and testing module is used to train and test the performance of truck travel category classification;

[0058] The instance verification module is used to apply the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments.

[0059] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements any one of the above-mentioned truck travel category classification methods based on multi-scale hypergraph convolutional deep clustering.

[0060] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements any one of the above-mentioned truck travel category classification methods based on multi-scale hypergraph convolutional deep clustering.

[0061] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements any one of the above-mentioned truck travel category classification methods based on multi-scale hypergraph convolutional deep clustering.

[0062] In the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering provided by the present invention, it is possible to rely on the fusion of hypergraph representation and hypergraph convolutional network technology to deeply model the implicit feature representations of truck individuals and their neighborhood group travel patterns, and on the basis of considering the "many-to-many" association of multi-scale spatial interactions of travel patterns, achieve accurate classification of truck groups, which is beneficial to improving the refined management level of managers for different category travel groups, so as to formulate classification policies and fine emergency response plans subsequently. Description of the Drawings

[0063] To more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0064] Figure 1 It is one of the schematic flowcharts of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention;

[0065] Figure 2 It is the second of the schematic flowcharts of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention;

[0066] Figure 3 It is the schematic structural diagram of the truck travel category division system based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention. Detailed implementation manners

[0067] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0068] Figure 1 It is one of the schematic flowcharts of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention.

[0069] Figure 2 It is the second of the schematic flowcharts of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention.

[0070] As Figure 1 and Figure 2 shown, this embodiment provides a truck travel category division method based on multi-scale hypergraph convolutional deep clustering, including:

[0071] Step 1, construct the truck travel feature matrices at each spatial scale. According to the historical travel record set of N trucks, obtain the historical travel record set of N trucks at a given time period; at the spatial scale s, extract F-dimensional travel features for each truck, and construct the travel feature matrix of N trucks Obtain the truck travel feature matrices X = {X1,..., X s ,..., X S}, where x s,i is the travel feature expression of the i-th truck at the spatial scale s, 1 ≤ i ≤ N; x i is the travel feature expression of the i-th truck, 1 ≤ i ≤ N; the F-dimensional travel features of each truck come from the time, space, and attribute dimensions respectively. The time dimension includes the number of times of fatigue driving of the truck driver, the night driving duration, the departure time of the truck, and the average daily travel duration; the space dimension includes the travel distance of the truck, the function of the truck stop point, and the average travel speed of the truck; the attribute dimension includes the license plate color and the number of transportation trips;

[0072] Step 2, construct the truck travel pattern hypergraph of each spatial scale, model the many-to-many high-order correlation relationship of trucks, and according to the travel feature matrix X of the trucks in Step 1 s , construct X s The travel pattern hypergraph G s =(V s , X s , W s , E s , H s ) at the spatial scale s. Take N trucks as N nodes on the graph to form a node set V s expressing all trucks; regard the F-dimensional travel features of each truck as the attributes of each node to form a node attribute set X s expressing the truck travel pattern; connect two adjacent nodes, and use the reciprocal of the Euclidean distance between them as the edge weight value of the corresponding nodes to form an edge weight value set W s expressing the importance degree of the neighborhood group; select the first k - 1 neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the i-th truck Analogize the above neighborhood group to a hyperedge to obtain a set of M hyperedges, forming a hyperedge set E s expressing the complex correlation relationship of the many-to-many travel patterns of the neighborhood group; H s is an N×N association matrix expressing the membership relationship between a single truck and the neighborhood group. The internal element h(v, e) of H s characterizes the membership relationship between node v and hyperedge e. The specific calculation method is shown in Equation (1):

[0073]

[0074] where v ∈ e means the node is on the hyperedge;

[0075] Step 3: Construct a multi-scale hypergraph convolutional deep clustering model for training and testing the classification of truck travel categories, including:

[0076] Step 3.1, use the autoencoder model to learn the implicit travel pattern representation of trucks;

[0077] Step 3.2: Using the hypergraph convolutional network structure, learn the implicit travel pattern representation of the truck neighborhood group in the local space;

[0078] Step 3.3: Using the cross self-attention mechanism, fuse the implicit travel pattern representations of the trucks and their neighborhood groups at various spatial scales, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale;

[0079] Step 3.4: Train and test the performance of truck travel category classification;

[0080] Step 4: Instance verification. Input the travel feature matrix X of N trucks at the (S + 1)-th spatial scale S+1 , construct the truck travel pattern hypergraph G' that fuses multi-scale spatial features S+1 , and achieve the classification of truck group travel categories; apply it to the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments.

[0081] The trucks in this embodiment can be divided into different groups based on the above steps. In implementation, they can be divided based on the types of goods carried by the trucks. For example, they can be divided into trucks transporting vegetables, trucks transporting fruits, trucks transporting garbage, etc. After classifying the trucks according to the goods they carry, it is helpful for subsequent refined management of different truck categories and facilitates the formulation of subsequent classification policies.

[0082] In practical applications, in step 1, information about the truck can be obtained from three dimensions: time, space, and attributes, and the category of the truck can be comprehensively judged from these three dimensions. Among them, the time information can include the departure time, travel duration, etc. For different goods, the departure time and travel duration may be different. For example, trucks transporting garbage may work from 5:00 to 6:00 in the morning or from 17:00 to 20:00 in the evening, while trucks transporting fruits and vegetables may need to work 24 hours a day to keep the goods fresh. The space information can include the stopping points of the truck and the travel speed, etc. Among them, the stopping points are generally the locations where the truck loads and unloads goods. For different goods, the loading and unloading locations of the truck must also be different. For example, trucks transporting fruits and vegetables may transport from the wholesale market to the vegetable market, or from the vegetable market to the supermarket, while trucks transporting garbage may transport from the residential area to the garbage station. In addition, the speed of the truck can also play an auxiliary role in judging the type of goods. For example, for safety reasons, the speed of trucks transporting garbage may be relatively slow, while for the freshness of the goods, the speed of trucks transporting fruits and vegetables may be relatively fast. In addition, for different trucks, the license plate color also limits the types of goods that can be transported. Based on this attribute information of the license plate color, it can also assist in classifying the truck categories.

[0083] The hypergraph constructed in step 2 is a generalized graph structure. Compared with the traditional graph structure where an edge can only connect two vertices, in the hypergraph structure, an edge can connect any number of vertices. The advantage of the hypergraph structure is that it can distinguish between hierarchical and non-hierarchical information structures, and any complex information structure can be established by these two types of information.

[0084] A neighborhood is a topological structure on a set. A neighborhood is a special curve. If a certain point is used as the center point, any open interval is called the neighborhood of that point.

[0085] The clustering in step 3 is a process of dividing a set of physical or abstract objects into similar object classes. A cluster is a set of data objects. These objects are similar to each other within the same cluster, but different from those in other clusters.

[0086] Specifically, in the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering provided in this embodiment, it is possible to rely on hypergraph representation and hypergraph convolutional network technologies to deeply model the implicit feature representation of the travel patterns of individual trucks and their neighborhood groups. On the basis of considering the many-to-many associations of multi-scale spatial interactions of travel patterns, accurate classification of truck groups can be achieved, which is beneficial to improving the refined management level of managers for different category travel groups, so as to formulate classification policies and fine emergency response plans subsequently.

[0087] In an exemplary embodiment, the use of the autoencoder model to learn the implicit travel pattern representation of trucks includes:

[0088] Introduce the autoencoder model, and through the built-in encoder, learn the truck travel feature matrix X at the spatial scale s in step 1 s The implicit travel pattern representation in the convolutional layer l (1 ≤ l ≤ L) Is calculated by formula (2):

[0089]

[0090] In the formula, σ ReLU Is a linear activation function; AutoEncoder is an autoencoder. When l = 1, Is the truck travel feature matrix X output in step 1 s ; W (l) Is the convolutional kernel weight of the l-th layer, and b (l) Is the bias value of the l-th layer. Both are weight parameters to be trained in the model;

[0091] Through the decoder built in the autoencoder, after L convolutional layers, the reconstructed Complete the decoding operation; minimize the average reconstruction error generated by the autoencoder model at all spatial scales Is expressed by formula (3) to obtain the truck travel pattern representation closest to the real situation:

[0092]

[0093] Among them, ||·|| Fro Is the F norm;

[0094] For the spatial scale s, according to the distribution of the truck travel feature matrix X s Obtain the probability distribution curve GT of truck travel categories in the real scenario; according to the distribution of the implicit feature representation of the truck travel pattern in the L-th convolutional layer Obtain the probability distribution curve TR of truck travel categories divided by the autoencoder, and measure the cumulative error between curve GT and curve TR at all spatial scales Is calculated by formula (4):

[0095]

[0096] Among them, GT sij And AE sij Are respectively the probabilities that truck i belongs to category j in curve GT and curve AE at the spatial scale s.

[0097] In an exemplary embodiment, the use of the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck neighborhood group in the local space includes:

[0098] Based on the truck travel pattern hypergraph G at any spatial scale s obtained in step 2 s , perform hypergraph convolutional network operations to capture the local spatial topological changes of the truck neighborhood group. For the l-th convolutional layer, the implicit travel pattern of the truck neighborhood group is calculated by formula (5):

[0099]

[0100] where Θ is the convolutional hyperparameter; when l = 1, is the truck travel feature matrix X obtained in step 1 s ; H s Tr is the transpose matrix of H s ; D s,v = [d(v)] is the degree matrix of node v at spatial scale s, which is used to describe the importance of node v in the hypergraph G s at the current spatial scale; the internal element d(v) is the sum of the weights of all hyperedges connected to this vertex v, and is calculated by formula (6):

[0101]

[0102] D s,e = [d(e)] is the degree matrix of hyperedge e at spatial scale s, which is used to describe the importance of hyperedge e in the hypergraph G s at the current spatial scale; the internal element d(e) is the sum of the number of all nodes connected by this hyperedge e, and is calculated by formula (7):

[0103]

[0104] In an exemplary embodiment, the use of the cross self-attention mechanism to fuse the implicit travel pattern representations of the aforementioned trucks and their neighborhood groups at each spatial scale, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale, includes:

[0105] Based on the implicit travel pattern representation of the truck in the l-th convolutional layer in step 3.1 and the implicit travel pattern representation of the truck neighborhood group in the l-th convolutional layer in step 3.2 Introduce the cross self-attention mechanism, and obtain the implicit travel pattern expression of the truck neighborhood group in the l-th convolutional layer at the s-th spatial scale through formula (8) Then, a new expression of the neighborhood group travel pattern of trucks in the l-th convolutional layer that fuses all spatial scales is obtained through formula (9).

[0106]

[0107]

[0108] In formula (8), CrossSelfAttention represents the cross self-attention mechanism; For the n-th vector sequence of; α mn is the m-th query and the attention weight of the n-th key indicating their correlation β mn is the m-th query and the attention weight of the n-th key indicating their correlation And is the attention scoring function, used to measure the correlation between the two in the formula, and is quantitatively expressed using the dot product model;

[0109] Based on the fused features of multiple spatial scales, update the corresponding hypergraphs at each spatial scale Correspondingly splice to obtain the dynamic hypergraph As the input of the hypergraph convolutional network in the (l + 1)-th convolutional layer, and cross self-attention fusion is performed with the output result corresponding to the (l + 1)-th layer of the autoencoder After that, perform the hypergraph convolutional network operation, and obtain the probability distribution result of the truck travel category in the (l + 1)-th convolutional layer through the following formula (10) Finally, complete the dynamic division of the truck travel category;

[0110]

[0111] In formula (10), W′ s 、H′ s 、D′ s,v 、D′ s,e are respectively the hyperedge weight set, incidence matrix, node degree matrix, and hyperedge degree matrix of the fused hypergraph of multiple spatial scales ;

[0112] For any spatial scale s, according to the distribution of the implicit travel pattern representation of trucks in the L-th convolutional layer obtain the probability distribution curve HC of the truck travel category obtained by hypergraph clustering at spatial scale s, and measure the cumulative error between curve HC and curve GT at all spatial scales Calculated by formula (11):

[0113]

[0114] Among them, HC sij refers to the probability that truck i belongs to category j in curve HC at spatial scale s.

[0115] In an exemplary embodiment, the training and testing of the performance of classifying truck travel categories includes:[[]]

[0116] From formula (12), based on the average reconstruction error and the and the to obtain the overall error

[0117]

[0118] Among them, η1, η2, and η3 are the weight coefficients of the three errors respectively;

[0119] Continuously loop and iterate the above steps 3.1 to 3.3 to minimize the cumulative error Until the value no longer changes, the model converges. At this time, the classified truck travel categories are truly available, and the training process of the model is completed;

[0120] Apply the trained model to the test dataset to verify the performance of the model in classifying truck travel categories at any spatial scale.

[0121] In an exemplary embodiment, the verification of the effective classification performance of the method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering includes:[[]]

[0122] Verify the effective classification performance of the method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering through four indicators: accuracy, normalized virtual information, average Rand index, and F1-score.

[0123] Among them, the full name of accuracy is (Accuracy, ACC), the full name of normalized virtual information is (Normalized Mutual Information, NMI), the full name of average Rand index is (Adjusted Rand Index, ARI), and the full name of F1-score is (F1-Score).

[0124] In practical applications, the larger the values of the four indicators, the better the classification performance. <H

[0125] In practical applications, the method for example verification of the truck travel category division method based on multi-scale hypergraph convolution deep clustering includes the following steps:

[0126] (1) Working environment

[0127] Use the PyTorch framework to write the model code. All experimental codes are compiled and executed on the Chinese Home Edition of Microsoft Windows 10, and run on a hardware workstation equipped with a "11th Gen intel(R) Core(TM) i5-11300H@3.10GHz" CPU and an "NVIDIA GeForce MX450 (16GB RAM)" GPU graphics card.

[0128] In this embodiment, three spatial scales are designed, namely the site scale, the block scale, and the administrative division.

[0129] (2) Introduction of the experimental dataset

[0130] The experimental datasets in this embodiment are the ACM dataset, the Cite dataset, and the self-collected truck GPS dataset. The ACM dataset contains 1870 sample dimensions, 3025 samples, and 3 categories; the Cite dataset contains 3703 sample dimensions, 3327 samples, and 6 categories; the self-collected truck GPS dataset contains 6-dimensional travel feature data of 154,119 trucks after data preprocessing from June 1, 2019 to June 15, 2019.

[0131] (3) Selection of the baseline model

[0132] In the comparison and ablation experiment sessions, 5 baseline models are selected to compare the performance with the proposed model on the given datasets.

[0133] k-means clustering model: A type of unsupervised learning method that discovers the hidden structure in data by dividing the observation points in the dataset into different groups or clusters;

[0134] AE (Autoencoder): The built-in encoder maps the input data to a low-dimensional latent space, and then performs clustering tasks in this space;

[0135] DBSCAN (Improved deep embedded clustering): A density-based clustering algorithm that can identify clusters of arbitrary shapes and is robust to outliers. This algorithm does not depend on the prior number of clusters and is suitable for exploring datasets with complex structures;

[0136] SDCN (Structured deep clustering network): Construct a k-nearest neighbor graph based on feature similarity, perform autoencoder pre-training, and capture the aggregation structure relationship between data;

[0137] AGCN (Attention-driven graph clustering network): Extract multi-scale node attributes and graph topology feature information, perform dynamic fusion, and execute the clustering task of data;

[0138] (4) Experimental parameter settings

[0139] In this embodiment, the "AdamW" optimizer is used to perform iterative training 200 times on a given dataset respectively to carry out the group division performance research experiment. In the experiment on the ACM dataset, the learning rate of all models is set to 10-3; in the experiment on the Cite dataset, the learning rate of all models is set to 10-4. In the experiments on the ACM dataset and the Cite dataset, in order to ensure fair comparison of all methods, the present invention optimizes and sets according to the network configuration parameters recommended by each baseline model. The network sizes of AE and AGCN are both set to "500-500-2000-10". In the experiment on the truck GPS dataset, for k-means, we calculate the silhouette coefficient for different numbers of clusters and find that the number of clusters 3 with the highest silhouette coefficient is the best number of clusters, and the silhouette coefficient value is output. DBSCAN finds the best parameter combination of eps = 1.1 and MinPts = 7 through a tuning strategy. The network size of the AE autoencoder is set to "500-500-2000-10". The network sizes of the built-in encoder and GCN module of SDCN are both set to "500-500-2000-10". For the method proposed in the present invention, the weight coefficients η1, η2, and η3 in formula (12) are respectively set to 1, 0.1, and 0.01.

[0140] (5) Selection of evaluation indicators

[0141] In the experiments on the ACM dataset and the Cite dataset, the clustering performance of each model is evaluated through four common indicators, namely accuracy (Accuracy, ACC), normalized mutual information (NMI), average Rand index (ARI), and F1-score (F1). Generally speaking, the larger the evaluation indicator, the better the effect.

[0142] In the experiment on the truck GPS dataset, the silhouette coefficient method is used to evaluate the clustering performance of each model. The larger the evaluation indicator, the better the effect.

[0143] (6) Experimental results and analysis

[0144] Table 1 shows the experimental results of group division comparison of each model on two public datasets, ACM and Cite. Generally speaking, the proposed method achieves the best clustering performance (see the bold data in Table 1). Compared with the baseline model SDCN, which has the second-best group division performance, the proposed method has achieved faster growth in various indicators on the two public datasets, namely: the ACC has increased by 1.62% and 2.47% respectively, the NMI has increased by 4.99% and 1.61% respectively, the ARI has increased by 3.86% and 2.86% respectively, and the F1 has increased by 1.63% and 0.36% respectively, which confirms that the present invention has good group division performance. This is attributed to the hypergraph module and feature fusion module proposed in the present invention. The former can fully learn the similarities of implicit travel patterns between individuals and between domain groups; the latter can effectively combine the travel pattern characteristics of individuals themselves and their neighboring groups to better describe the implicit clustering characteristics of target individuals, thereby improving the performance of group division.

[0145] Table 1 Comparative experimental results of each model on public datasets

[0146]

[0147] Table 2 reflects the improvement of the hypergraph module and feature fusion module of the proposed method on the group division performance. The experimental results in the second row of each dataset in the table are the distribution of the group division performance when the proposed method only carries the first module; the experimental results in the third row of each dataset in the table are the distribution of the group division performance when the proposed method carries the above two modules. It can be seen from the table that adding the hypergraph module can make the ACC of the proposed model increase by 0.4% and 2.72% respectively in the two datasets, the NMI increase by 0.23% and 2.9% respectively, the ARI increase by 3.22% and 3.15% respectively, and the F1 index decreases, indicating that only constructing the hypergraph module may not be sensitive on public datasets, while other indexes have increased, which confirms that the hypergraph module can effectively explain the complex association characteristics of many-to-many between individuals; after adding the feature fusion module, on the basis of the hypergraph module, the proposed model can make the ACC increase by 0.43% and 0.8% respectively again, the NMI increase by 5.07% and 2.35% respectively, the ARI increase by 6.08% and 3.07% respectively, and the F1 increase by 2.61% and 2.52% respectively. This fully shows that the design of the hypergraph module and feature fusion module in the present invention helps to improve the accurate division performance of groups.

[0148] Table 2 Ablation experimental results of the built-in modules of the proposed method on public datasets

[0149]

[0150] Table 3 reflects the effect of the proposed method in the field of trucks for group division. Compared with the traditional algorithms K-means and DBSCAN, the effect of group division of the proposed method in this paper has increased by about 8.91% and 4.68% respectively; compared with the algorithms AE, SDCN, and AGCN that introduce deep learning technology, the effect of group division has increased by about 8.14%, about 4.00%, and about 3.02% respectively. It can be seen that the proposed method in this paper can not only effectively process the high-order relationships in truck travel data, but also provide a more accurate group division effect than traditional clustering algorithms and other deep learning methods, thus more accurately dividing truck travel categories.

[0151] Table 3 Classification effect of the travel pattern categories of trucks by the proposed method on the self-collected truck GPS dataset

[0152]

[0153] The truck travel category division system based on multi-scale hypergraph convolutional deep clustering provided by the present invention will be described below. The truck travel category division system based on multi-scale hypergraph convolutional deep clustering described below can be correspondingly referred to the truck travel category division method based on multi-scale hypergraph convolutional deep clustering described above.

[0154] Figure 3 It is a schematic structural diagram of the truck travel category division system based on multi-scale hypergraph convolutional deep clustering provided by an embodiment of the present invention.

[0155] As Figure 3 shown, the truck travel category division system based on multi-scale hypergraph convolutional deep clustering provided by this embodiment includes:

[0156] A truck travel feature matrix construction module 301, configured to obtain the historical travel record set of N trucks in a given time period according to the historical travel record set of N trucks; at the spatial scale s, extract F-dimensional travel features for each truck, and construct a travel feature matrix X of N trucks s ;

[0157] A truck travel pattern hypergraph construction module 302, configured to construct a travel pattern hypergraph G s of X s at the spatial scale s, where the N trucks are used as N nodes on the graph, forming a node set V s = (V s , X s , W s , E s , H s ), taking the N trucks as N nodes on the graph, forming a node set V s that represents all trucks; regarding the F-dimensional travel features of each truck as the attributes of each node, forming a node attribute set X that represents the truck travel patterns ; Connect adjacent pairwise nodes, and use the reciprocal of their Euclidean distance as the edge weight value of the corresponding nodes to form an edge weight value set \(W\) that expresses the importance degree of the neighborhood group. s ; Select the top \(k - 1\) neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the \(i\)-th truck. Analogy the above neighborhood group to a hyperedge to obtain a set of \(M\) hyperedges, forming a hyperedge set \(E\) that expresses the complex association relationship of the many-to-many travel patterns of the neighborhood group. s , \(H\) s is an \(N\times N\) association matrix that expresses the membership relationship between a single truck and the neighborhood group;

[0158] The implicit travel pattern representation learning module 303 of the truck is used to utilize the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck's neighborhood group in the local space;

[0159] The neighborhood group implicit travel pattern representation learning module 304 is used to utilize the cross self-attention mechanism to fuse the implicit travel pattern representations of the truck and the truck's neighborhood group at each spatial scale, and perform deep clustering through the hypergraph convolutional network to obtain the division results of the truck travel categories at each spatial scale;

[0160] The hypergraph convolutional deep clustering module 305 is used to utilize the cross self-attention mechanism to fuse the implicit travel pattern representations of the truck and the truck's neighborhood group at each spatial scale, and perform deep clustering through the hypergraph convolutional network to obtain the division results of the truck travel categories at each spatial scale;

[0161] The training and testing module 306 is used to train and test the performance of the truck travel category division;

[0162] The instance verification module 307 is used to apply the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective division performance of the truck travel category division method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments.

[0163] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering, characterized in that including: Step 1: Construct the truck travel feature matrix at each spatial scale. Based on the historical travel record set of N trucks, obtain the historical travel record set of N trucks in a given time period; at spatial scale s, extract F-dimensional travel features for each truck and construct the travel feature matrix of N trucks. Obtain the truck travel feature matrices X = {X1, …, X s , …, X S} at S hierarchical spatial scales, where x s,i is the travel feature expression of the i-th truck at spatial scale s, 1 ≤ i ≤ N; x i is the travel feature expression of the i-th truck, 1 ≤ i ≤ N; the F-dimensional travel features of each truck come from the time, space, and attribute dimensions respectively. The time dimension includes the number of times of truck driver fatigue driving, the night driving duration, the truck departure time, and the average daily travel duration; the space dimension includes the truck travel distance, the function of the truck stop point, and the average truck travel speed; the attribute dimension includes the license plate color and the number of transportation trips. Step 2: Construct the hypergraph of truck travel patterns at each spatial scale, model the high-order association relationship of many-to-many trucks, and construct the travel feature matrix X of the trucks according to the description in Step 1 s , construct X s The travel pattern hypergraph G s =(V s , X s , W s , E s , H s ) at the spatial scale s. Consider N trucks as N nodes on the graph to form a node set V s that represents all trucks; regard the F-dimensional travel features of each truck as the attributes of each node to form a node attribute set X s that represents the truck travel pattern; connect adjacent pairs of nodes, and use the reciprocal of their Euclidean distance as the edge weight value of the corresponding nodes to form an edge weight value set W s that represents the importance degree of the neighborhood group; select the top k - 1 neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the i-th truck Analogize the above neighborhood group to a hyperedge to obtain a set of M hyperedges, and form a hyperedge set E s that represents the complex association relationship of the many-to-many travel pattern of the neighborhood group; H s is an N×N association matrix that represents the membership relationship between a single truck and the neighborhood group. The internal element h(v, e) of H s characterizes the membership relationship between node v and hyperedge e. The specific calculation method is shown in Equation (1): where \(v\in e\) indicates that the node is on the hyperedge; Step 3: Construct a multi-scale hypergraph convolutional deep clustering model for training and testing the classification of truck travel categories, including: Step 3.1: Use an autoencoder model to learn the implicit travel pattern representation of trucks; Step 3.2: Use the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck neighborhood group in the local space; Step 3.3: Use the cross self-attention mechanism to fuse the implicit travel pattern representations of trucks and their neighborhood groups at various spatial scales, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale; Step 3.4: Train and test the performance of truck travel category classification; Step 4: Instance verification. Input the travel feature matrix X of N trucks at the (S + 1)-th spatial scale S+1 , and construct a hypergraph G' of truck travel patterns that integrates multi-scale spatial features S+1 , and realize the classification of truck group travel categories; apply it to the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments.

2. The method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering according to claim 1, wherein The use of the autoencoder model to learn the implicit travel pattern representation of trucks includes: Introduce an autoencoder model. Through the built-in encoder, learn the truck travel feature matrix X at the spatial scale s in step 1 s The implicit travel pattern representation in the convolutional layer l where 1 ≤ l ≤ L, 1 ≤ s ≤ S, and it is calculated by formula (2): where, σ ReLU is a linear activation function; AutoEncoder is an autoencoder. When l = 1, is the truck travel feature matrix X output in Step 1 s ; W (l) is the convolutional kernel weight of the l-th layer, and b (l) is the bias value of the l-th layer. Both are weight parameters to be trained in the model; Through the decoder built into the autoencoder, after passing through L convolutional layers, the reconstructed completes the decoding operation; minimizing the average reconstruction error generated by the autoencoder model at all spatial scales expressed by formula (3) to obtain a representation of the truck travel pattern that is closest to the real situation: where, ||·|| Fro is the Frobenius norm; For the spatial scale s, according to the truck travel characteristic matrix X s 's distribution, obtain the probability distribution curve GT of truck travel categories in the real scenario; according to the distribution of the implicit features of the truck travel pattern in the L-th convolutional layer 's distribution, obtain the probability distribution curve TR of truck travel categories divided by the autoencoder, and measure the cumulative error between curve GT and curve TR at all spatial scales Calculated by formula (4): Among them, GT sij and TR sij are the probabilities that truck i belongs to category j in curve GT and curve TR at spatial scale s, respectively.

3. The method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering according to claim 1, characterized in that The use of the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck neighborhood group in the local space includes: According to the truck travel pattern hypergraph \(G\) at any spatial scale \(s\) obtained in step 2 s , perform hypergraph convolutional network operations to capture the local spatial topological changes of the truck neighborhood group. For the \(l\)-th convolutional layer, the implicit travel pattern of the truck neighborhood group is calculated by formula (5): where Θ is the convolution hyperparameter; when l = 1, is the truck travel feature matrix X obtained in step 1 s ; H s Tr is the transpose matrix of H s ; D s,v = [d(v)] is the degree matrix of node v at spatial scale s, which is used to describe the importance of node v in the hypergraph G s ; the internal element d(v) is the sum of the weights of all hyperedges connected to this vertex v, which is calculated by formula (6): D s,e = [d(e)] is the degree matrix of the hyperedge e at the spatial scale s, which is used to describe the importance of the hyperedge e in the hypergraph G s at the current spatial scale; the internal element d(e) is the sum of the number of all nodes connected by the hyperedge e, which is calculated by formula (7):

4. The method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering according to claim 2, characterized in that, The use of the cross self-attention mechanism to fuse the implicit travel pattern representations of the aforementioned trucks and their neighborhood groups at various spatial scales, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale includes: The implicit travel pattern representation of trucks in the l-th convolutional layer according to Step 3.1 The implicit travel pattern representation of the truck neighborhood group in the l-th convolutional layer in Step 3.2 Introduce the cross self-attention mechanism, and obtain the implicit travel pattern expression of the truck neighborhood group in the l-th convolutional layer at the s-th spatial scale through formula (8) Then, obtain the new expression of the travel pattern of the truck neighborhood group in the l-th convolutional layer that fuses all spatial scales through formula (9) In formula (8), CrossSelfAttention represents the cross self-attention mechanism; is the nth vector sequence of mn α is the attention weight of the mth query and the nth key β mn is the attention weight of the mth query and the nth key indicating their correlation and is the attention scoring function, which is used to measure the correlation between the two in the formula and is quantitatively expressed using the dot product model; Update the corresponding hypergraphs at each spatial scale based on the fusion features across multiple spatial scales Perform corresponding splicing to obtain the dynamic hypergraph As the input to the hypergraph convolutional network at convolutional layer l + 1, and the corresponding output result of the autoencoder at layer l + 1 After performing cross - self - attention fusion, perform hypergraph convolutional network operations, and obtain the probability distribution result of the truck travel category at the (l + 1)-th convolutional layer through the following formula (10) Finally, complete the dynamic classification of truck travel categories; In formula (10), W' s , H' s , D' s,v , D' s,e are respectively the hyper-edge weight set, the incidence matrix, the node degree matrix, and the hyper-edge degree matrix of multiple spatially-scaled fusion hypergraphs ; For any spatial scale s, according to the distribution of the representation of the hidden travel patterns of trucks in the L-th convolutional layer obtain the probability distribution curve HC of truck travel categories obtained by hypergraph clustering at spatial scale s, and measure the cumulative error between curve HC and curve GT at all spatial scales which is calculated by formula (11): Among them, HC sij represents the probability that truck i belongs to category j in curve HC at spatial scale s.

5. The method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering according to claim 4, wherein The training and testing of the performance of truck travel category classification includes: Based on the average reconstruction error from Equation (12) and the and the the overall error is obtained where \(\eta_1\), \(\eta_2\), and \(\eta_3\) are the weight coefficients of the three errors respectively; Continuously iterate the above steps 3.1 to 3.3 to minimize the cumulative error until the value no longer changes, the model converges, and at this time, the truck travel categories obtained by the division are truly available, completing the training process of the model; Apply the trained model to the test dataset to verify the performance of the model in classifying truck travel categories at any spatial scale.

6. The method for classifying truck travel categories based on multi-scale hypergraph convolutional deep clustering according to claim 1, characterized in that, The verification of the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments includes: Verify the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through four indicators: accuracy, normalized virtual information, average Rand index, and F1-score.

7. A truck travel category classification system based on multi-scale hypergraph convolutional deep clustering, characterized in that including: The truck travel feature matrix construction module is used to obtain the historical travel record set of N trucks in a given time period according to the historical travel record set of N trucks; at the spatial scale s, F-dimensional travel features are extracted for each truck, and the travel feature matrix X of N trucks is constructed s ; Truck travel pattern hypergraph construction module, which is used to construct a travel pattern hypergraph \(G\) of \(X\) at the spatial scale \(s\) according to the travel feature matrix \(X\) of the truck s , construct \(X\) s in the travel pattern hypergraph \(G\) of \(X\) at the spatial scale \(s\) s =(V s , \(X\) s , \(W\) s , \(E\) s , \(H\) s ). Regarding \(N\) trucks as \(N\) nodes on the graph, a node set \(V\) representing all trucks is formed s ; regarding the \(F\)-dimensional travel features of each truck as the attributes of each node, a node attribute set \(X\) representing the travel patterns of trucks is formed s ; connecting adjacent pairs of nodes, and using the reciprocal of the Euclidean distance between the two as the edge weight value of the corresponding nodes, an edge weight value set \(W\) representing the importance degree of the neighborhood group is formed s ; selecting the first \(k - 1\) neighborhood trucks with the smallest Euclidean distance to form the neighborhood group of the \(i\)-th truck Regarding the above neighborhood group as a hyperedge, an \(M\)-hyperedge set is obtained, forming a hyperedge set \(E\) representing the complex association relationship of the many-to-many travel patterns of the neighborhood group s , \(H\) s is an \(N\times N\) association matrix representing the membership relationship between a single truck and the neighborhood group; A module for learning the implicit travel pattern representation of trucks, which is used to use the hypergraph convolutional network structure to learn the implicit travel pattern representation of the truck neighborhood group in the local space; A module for learning the implicit travel pattern representation of the neighborhood group, which is used to use the cross self-attention mechanism to fuse the implicit travel pattern representations of trucks and their neighborhood groups at various spatial scales, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale; A hypergraph convolutional deep clustering module, which is used to use the cross self-attention mechanism to fuse the implicit travel pattern representations of trucks and their neighborhood groups at various spatial scales, and perform deep clustering through the hypergraph convolutional network to obtain the classification results of truck travel categories at each spatial scale; A training and testing module, which is used to train and test the performance of truck travel category classification; An example verification module, which is used to apply the ACM dataset, Cite dataset, and truck GPS dataset, and verify the effective classification performance of the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering through comparison and ablation experiments.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the truck travel category classification method based on multi-scale hypergraph convolutional deep clustering as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a processor, it implements the truck trip category division method based on multi-scale hypergraph convolutional deep clustering as described in any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the truck trip category division method based on multi-scale hypergraph convolutional deep clustering as described in any one of claims 1-6.

Images