A Method and System for Constructing Trajectory Archives Based on Temporal Dependency Graphs and Neural Networks

Abstract

The application discloses a trajectory archive construction method and system based on a time sequence dependent graph and a neural network, and particularly relates to the technical field of data processing and graph neural networks, and is used for solving the problem that the quality of a time sequence dependent graph constructed due to sparse and incomplete trajectory data is poor, thereby affecting the accuracy of a trajectory archive; an initial time sequence dependent graph is constructed by analyzing the space-time proximity relationship of original trajectory data, then an isolation forest algorithm is used to calculate the individual mode deviation of each edge in the graph, abnormal edges are identified based on a deviation threshold, the moving intention consistency of the connected trajectory segments is analyzed to deduce intention coherence, semantic consistency verification is performed in combination with the node region function attribute similarity and intention coherence to identify semantic conflict edges, finally, a multi-path replaceability decision mechanism is used to operate the semantic conflict edges to optimize the graph structure, and a weighted graph neural network is used to process the optimized time sequence dependent graph to generate a trajectory archive vector.

Description

Technical Field

[0001] This invention relates to the fields of data processing and graph neural network technology, and more specifically, to a method and system for constructing trajectory archives based on temporal dependency graphs and neural networks. Background Technology

[0002] The widespread adoption of mobile sensing technologies has enabled the recording and analysis of massive amounts of spatiotemporal trajectory data. To extract high-level behavioral patterns and semantic information from this low-level trajectory data, trajectory profile construction technology has become a key support for fields such as smart cities, transportation planning, and location services. Currently, methods combining temporal dependency graphs and neural networks represent the mainstream technological direction in this field. These methods typically first convert the original trajectory sequence into an explicit temporal dependency graph structure, where nodes represent key locations or road segments, and edges represent the spatiotemporal transition relationships of the moving object. Then, a graph neural network is used to learn the representation of the constructed graph, ultimately generating a compact profile vector that characterizes the behavioral features of the moving object. This method, by explicitly modeling the complex dependencies between trajectory points using a graph structure and leveraging the powerful representational capabilities of neural networks, can effectively capture spatiotemporal patterns in mobile behavior.

[0003] However, the effectiveness of existing methods based on temporal dependency graphs and neural networks largely depends on the quality and accuracy of the constructed temporal dependency graph. In practical applications, due to factors such as signal obstruction, device power consumption limitations, and packet loss, raw trajectory data generally exhibits high sparsity and incompleteness, resulting in graph structures constructed from such data containing a large number of unreliable connections. The core learning mechanism of graph neural networks is based on the isomorphism assumption of "semantic similarity between connected nodes," treating all connections in the graph with equal message aggregation. When this assumption is violated due to data sparsity, the information propagation process of the neural network confuses the semantics of originally unrelated nodes, leading to inherent biases in the final generated trajectory profile, making it difficult to accurately and reliably represent the true behavior patterns of moving objects. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for constructing trajectory files based on temporal dependency graphs and neural networks to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Trajectory archive construction methods based on temporal dependency graphs and neural networks include:

[0007] S1. By analyzing the spatiotemporal proximity relationships in the original trajectory data of the moving object, a temporal dependency graph containing nodes and transition edges is generated.

[0008] S2. By using the isolated forest algorithm to analyze the degree of conformity between the current trajectory transfer characteristics and the historical trajectory pattern characteristics of the moving object, the individual pattern deviation of each edge in the time-series dependency graph is generated.

[0009] S3. By analyzing the movement intention of trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold. Figure 1 Consistency, inferring the continuity of the intention of the moving object;

[0010] S4. By analyzing the functional attribute similarity of the regions where the nodes at both ends of the edge are located and combining it with the intent coherence, a semantic consistency verification result including semantically conflicting edges is generated.

[0011] S5. By analyzing the multi-path substitutability decision of semantic conflict edges in the temporal dependency graph, the retention, weight reduction or deletion operations of semantic conflict edges are performed to generate a temporal dependency graph with optimized structure.

[0012] S6. By analyzing the processing results of the weighted graph neural network on the optimized temporal dependency graph, the trajectory archive vector of the moving object is generated.

[0013] Furthermore, by analyzing the spatiotemporal proximity relationships in the original trajectory data of the moving object, a temporal dependency graph containing nodes and transition edges is generated, including:

[0014] The original trajectory data of the moving object is clustered into trajectory points, and the trajectory points that are spatially close are clustered to generate key location nodes;

[0015] Based on the time sequence of trajectory points, analyze the transition relationship between key location nodes of adjacent trajectory points and generate transition edges between key location nodes;

[0016] Construct a temporal dependency graph using key location nodes and transition edges.

[0017] Furthermore, by employing the isolated forest algorithm to analyze the degree of consistency between the current trajectory transition characteristics and the historical trajectory pattern characteristics of the moving object, the individual pattern deviation of each edge in the temporal dependency graph is generated, including:

[0018] Extract trajectory transition features corresponding to transition edges in the temporal dependency graph. The trajectory transition features include displacement, duration, and region type.

[0019] Obtain trajectory data of the moving object within a historical time period, and construct a set of historical trajectory pattern features of the moving object based on the historical trajectory data;

[0020] The Isolation Forest algorithm is used to analyze the degree of conformity between the current trajectory transfer features and the features in the historical trajectory pattern feature set of the moving object, and to calculate the individual pattern deviation of each edge in the generated temporal dependency graph.

[0021] Furthermore, the Isolation Forest algorithm is used to analyze the degree of conformity between the current trajectory transition features and the features in the historical trajectory pattern feature set of the moving object, and to calculate the individual pattern deviation of each edge in the generated time-series dependency graph. This includes: inputting the historical trajectory pattern feature set of the moving object into the Isolation Forest algorithm for model training; inputting the current trajectory transition features corresponding to each edge in the time-series dependency graph into the trained Isolation Forest algorithm model; and calculating the anomaly score of the current trajectory transition features through the Isolation Forest algorithm model to generate the individual pattern deviation of each edge in the time-series dependency graph.

[0022] Furthermore, by analyzing the movement intention of trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold... Figure 1 Consistency, inferring the coherence of the intention of the moving object, including:

[0023] Identify edges in a temporal dependency graph where the deviation of an individual pattern exceeds a preset threshold.

[0024] Extract the trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold;

[0025] The movement pattern features of the moving objects in the extracted trajectory segments are analyzed. The movement pattern features include velocity change sequences and direction change sequences.

[0026] Analysis of the movement intention of trajectory segments based on movement pattern characteristics Figure 1 Consistency, inferring the coherence of the moving object's intention: Calculate the statistical consistency index of the velocity change sequence of the trajectory segment, calculate the statistical consistency index of the direction change sequence of the trajectory segment, and comprehensively analyze the movement intention of the trajectory segment based on the statistical consistency indices of the velocity change sequence and the direction change sequence. Figure 1 Consistency, based on the intention to move Figure 1 The consistency analysis results lead to the deduction of the intentional coherence of the moving object.

[0027] Furthermore, by analyzing the functional attribute similarity of the regions where the nodes at both ends of an edge are located and combining it with intent coherence, a semantic consistency verification result including semantically conflicting edges is generated, including:

[0028] Obtain the functional attribute information of the regions where the nodes at both ends of the edge in the time-series dependency graph are located;

[0029] Analyze the similarity between the functional attribute information of the regions where the nodes at both ends of the edge are located;

[0030] Based on the inferred intent coherence of the moving objects, analyze the consistency between the degree of similarity of functional attributes and intent coherence;

[0031] Based on the consistency analysis results of functional attribute similarity and intent coherence, a semantic consistency verification result including semantic conflict edges is generated.

[0032] Furthermore, by analyzing the retention, weight reduction, or deletion operations of semantically conflicting edges in the multi-path substitutability decision of the temporal dependency graph, a structurally optimized temporal dependency graph is generated, including:

[0033] Obtain the semantically conflicting edges contained in the semantic consistency verification result;

[0034] For each semantically conflicting edge, search the temporal dependency graph for a set of alternative paths connecting the nodes at both ends of the corresponding semantically conflicting edge;

[0035] Analyze the reliability of paths in the alternative path set and evaluate the multi-path substitutability of semantically conflicting edges;

[0036] Based on the evaluation results of the substitutability of multiple paths, the decision is made to perform retention, weight reduction or deletion operations on semantically conflicting edges: if there are multiple highly reliable alternative paths for a semantically conflicting edge, the edge is deleted; if there are no alternative paths for a semantically conflicting edge, the edge is retained but its weight is reduced; if there are a few low-reliability alternative paths for a semantically conflicting edge, the edge is weighted according to the comprehensive evaluation results of the reliability of alternative paths.

[0037] Based on the operation results of semantically conflicting edges, a temporal dependency graph with optimized structure is generated.

[0038] Furthermore, by analyzing the processing results of the weighted graph neural network on the structure-optimized temporal dependency graph, trajectory archive vectors of moving objects are generated, including:

[0039] The optimized temporal dependency graph is input into a weighted graph neural network.

[0040] The neighborhood feature information of nodes in the time-dependent graph is aggregated through a weighted graph neural network structure optimization.

[0041] Based on the aggregation results of node neighborhood feature information by the weighted graph neural network, a node representation vector is generated;

[0042] Aggregate the node representation vectors of all nodes in the time-dependent graph to generate the trajectory archive vector of the moving object.

[0043] Furthermore, the neighborhood feature information of nodes in the optimized temporal dependency graph is aggregated through a weighted graph neural network, including: calculating the weights of neighboring nodes based on the weights of edges in the optimized temporal dependency graph; weighting and summing the feature vectors of each neighboring node of the target node according to the corresponding neighboring node weights; and performing linear combination and nonlinear transformation on the weighted summation result and the target node's own feature vector to generate a new feature vector after aggregation of the target node.

[0044] On the other hand, the present invention provides a trajectory file construction system based on temporal dependency graphs and neural networks, comprising:

[0045] The proximity analysis module is used to generate a temporal dependency graph containing nodes and transition edges by analyzing the spatiotemporal proximity relationships in the original trajectory data of moving objects.

[0046] The deviation generation module is used to analyze the degree of conformity between the current trajectory transfer characteristics and the historical trajectory pattern characteristics of the moving object by using the isolated forest algorithm, and generate the individual pattern deviation of each edge in the time-series dependency graph.

[0047] The coherent deduction module is used to analyze the movement intention of trajectory segments connected by edges whose deviation from individual patterns exceeds a preset threshold. Figure 1 Consistency, inferring the continuity of the intention of the moving object;

[0048] The semantic verification module is used to generate semantic consistency verification results, including semantically conflicting edges, by analyzing the functional attribute similarity of the regions where the nodes at both ends of the edge are located and combining it with intent coherence.

[0049] The optimization processing module is used to generate a structurally optimized temporal dependency graph by analyzing the multi-path substitutability decisions of semantically conflicting edges in the temporal dependency graph and performing operations such as retention, weight reduction or deletion of semantically conflicting edges.

[0050] The archive generation module is used to generate trajectory archive vectors of moving objects by analyzing the processing results of the time-series dependency graph after structural optimization by a weighted graph neural network.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. By constructing a temporal dependency graph and performing multi-level structural optimization and semantic verification, the accuracy and reliability of the final generated trajectory archive vectors can be significantly improved. Key location nodes are generated through trajectory point clustering, and a temporal dependency graph is constructed based on spatiotemporal transition relationships, providing a stable and reliable basic graph structure for subsequent analysis. Then, individual pattern deviation calculation is introduced, and the isolated forest algorithm is used to effectively identify abnormal transition edges. Finally, the movement intention... Figure 1 Consistency analysis and functional attribute similarity assessment enabled accurate identification of semantically conflicting edges in the graph, ensuring that the temporal dependency graph was fully purified both structurally and semantically, providing high-quality input for subsequent neural network processing.

[0053] 2. Through a multi-path substitutability decision-making mechanism, semantically conflicting edges can be intelligently retained, downweighted, or deleted, maintaining graph connectivity while eliminating interference from noisy edges. Finally, the node features in the graph are aggregated and optimized through a weighted graph neural network, and the generated trajectory profile vector can more accurately represent the movement patterns and behavioral characteristics of moving objects. From raw trajectory data to high-quality trajectory profile vectors, each link is closely connected and mutually supportive, effectively solving the trajectory profile deviation problem caused by data sparsity and incompleteness, and enhancing the practical value of trajectory profiles in various downstream applications. Attached Figure Description

[0054] Figure 1 This is a flowchart of the trajectory file construction method based on temporal dependency graphs and neural networks of the present invention;

[0055] Figure 2 This is a schematic diagram of the trajectory file construction system based on temporal dependency graphs and neural networks of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1: Figure 1 The present invention provides a trajectory archive construction method based on temporal dependency graphs and neural networks, including:

[0058] S1. By analyzing the spatiotemporal proximity relationships in the original trajectory data of the moving object, a temporal dependency graph containing nodes and transition edges is generated.

[0059] S2. By using the isolated forest algorithm to analyze the degree of conformity between the current trajectory transfer characteristics and the historical trajectory pattern characteristics of the moving object, the individual pattern deviation of each edge in the time-series dependency graph is generated.

[0060] S3. By analyzing the movement intention of trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold. Figure 1 Consistency, inferring the continuity of the intention of the moving object;

[0061] S4. By analyzing the functional attribute similarity of the regions where the nodes at both ends of the edge are located and combining it with the intent coherence, a semantic consistency verification result including semantically conflicting edges is generated.

[0062] S5. By analyzing the multi-path substitutability decision of semantic conflict edges in the temporal dependency graph, the retention, weight reduction or deletion operations of semantic conflict edges are performed to generate a temporal dependency graph with optimized structure.

[0063] S6. By analyzing the processing results of the weighted graph neural network on the optimized temporal dependency graph, the trajectory archive vector of the moving object is generated.

[0064] S1. By analyzing the spatiotemporal proximity relationships in the original trajectory data of the moving object, a temporal dependency graph containing nodes and transition edges is generated. Specific implementation includes:

[0065] In constructing the temporal dependency graph, the original trajectory data of the moving object is first subjected to trajectory point clustering. This clustering process employs a density-based spatial clustering method, specifically determining spatial proximity by calculating the Euclidean distance between trajectory points. A neighborhood radius parameter is set to a value within the range of 50 to 200 meters. This parameter range is determined based on the density of the urban road network and the positioning accuracy; for example, a smaller radius value, such as 50 meters, can be used in densely populated urban areas, while a larger radius value, such as 200 meters, can be used in suburban areas. Simultaneously, a minimum number of points is set, for example, from 5 to 15 trajectory points, ensuring the statistical significance of the clustering results. Through this clustering process, spatially adjacent trajectory points are aggregated to generate key location nodes. Each key location node represents a frequently visited geographical area of ​​the moving object, and its location is determined by the latitude and longitude coordinates of the cluster center point, which is calculated using the arithmetic mean of the latitude and longitude coordinates of all member trajectory points.

[0066] After generating key location nodes, the transition relationships between adjacent key location nodes are analyzed based on the time sequence of the trajectory points. Specifically, all trajectory points are traversed in timestamp order. When two consecutive trajectory points belong to different key location nodes, a transition relationship is recorded between these two key location nodes. The number of transitions between each pair of key location nodes in all trajectory data is counted. When the number of transitions reaches, for example, 3 or more, a transition edge is generated between this pair of key location nodes. Each transition edge includes the number of transitions, average transition time, and most recent transition timestamp attributes. The average transition time is obtained by averaging the time differences of all adjacent trajectory points, with the time difference in seconds, and the result converted to minutes.

[0067] Finally, a temporal dependency graph is constructed using key location nodes and transition edges. The temporal dependency graph is represented by a directed graph structure, where key location nodes are the vertices and transition edges are the directed edges connecting the vertices. Each key location node includes a node identifier, latitude and longitude coordinates, and a visit frequency attribute, where the visit frequency is obtained by counting the total number of trajectory points falling within the node's range. Each transition edge includes a source node identifier, a target node identifier, and a transition weight attribute, where the transition weight is dynamically calculated based on the number of transitions and the transition time. The calculation formula is the ratio of the number of transitions to the average transition time, reflecting a comprehensive indicator of transition frequency and efficiency. The constructed temporal dependency graph is stored in computer memory in the form of an adjacency list, facilitating subsequent graph traversal and feature extraction operations. The adjacency list data structure includes a node list and an edge list. The node list records the attribute information of all key location nodes, and the edge list records the attributes and connection relationships of each transition edge. A node index is also established to improve query efficiency.

[0068] The construction of the temporal dependency graph also includes data quality control steps: preprocessing the trajectory data to remove obviously abnormal trajectory points, such as those with speeds exceeding 150 km / h; post-processing the clustering results to merge key location nodes that are too close, such as merging node pairs less than 20 meters apart; and filtering transition edges to remove edges with too few transitions, such as edges with fewer than 3 transitions. These quality control measures ensure that the final constructed temporal dependency graph accurately reflects the actual movement patterns of moving objects, providing a reliable foundation for subsequent analysis. The entire construction process is automated using a computer program written in Python, utilizing the scikit-learn library for clustering calculations and the NetworkX library for graph structure construction and storage.

[0069] S2. By employing the Isolation Forest algorithm, the degree of conformity between the current trajectory transition characteristics and the historical trajectory pattern characteristics of the moving object is analyzed, generating the individual pattern deviation of each edge in the time-series dependency graph. Specific implementation includes:

[0070] In generating the individual pattern deviation of each edge in the temporal dependency graph, the trajectory transfer features corresponding to the transfer edges in the graph are first extracted. These features include three dimensions: displacement, duration, and region type. Displacement is calculated by determining the Euclidean distance between the two key nodes connected by the transfer edge, in meters. Duration is calculated by determining the time difference between the transfers between the two key nodes, in seconds. Region type is obtained by acquiring the interest point category information of the regions where the two key nodes are located, such as commercial areas, residential areas, or industrial areas. The extraction of these features is based on the existing node location information and edge transfer attributes in the temporal dependency graph, ensuring the accuracy and consistency of the data source. During feature extraction, a coordinate transformation formula is used to convert latitude and longitude coordinates to planar coordinates to accurately calculate the Euclidean distance between nodes.

[0071] The process involves acquiring trajectory data of a moving object within a historical timeframe and constructing a historical trajectory pattern feature set based on this data. The historical timeframe is typically set to, for example, a range of 30 to 90 days, sufficient to cover the object's regular activity patterns. All transfer feature samples are extracted from the historical trajectory data, each sample containing three feature dimensions: displacement, duration, and region type. The extracted feature samples undergo data cleaning to remove obvious outliers, such as abnormal transfer records with displacements exceeding 50 kilometers or durations exceeding 24 hours. The 3σ principle is used to identify outliers during the cleaning process: the mean and standard deviation of each feature dimension are calculated, and values ​​exceeding the mean ± three times the standard deviation are considered outliers. The cleaned feature samples form the historical trajectory pattern feature set of the moving object, which characterizes the object's normal behavior patterns.

[0072] The Isolation Forest algorithm is used to analyze the degree of consistency between the current trajectory transition features and the features in the historical trajectory pattern feature set of the moving object. First, the historical trajectory pattern feature set of the moving object is input into the Isolation Forest algorithm for model training. The number of trees in the Isolation Forest algorithm is set to, for example, 100, and the sampling subset size of each tree is, for example, 256 samples. The algorithm constructs multiple isolation trees by randomly selecting features and random split values. Normal samples require more random splits to be isolated, while abnormal samples are more easily isolated. During training, each isolation tree recursively selects features and split values ​​randomly until every sample is isolated, forming a complete isolation tree structure. After training, the Isolation Forest algorithm model is obtained, which learns the data distribution characteristics of the historical trajectory pattern feature set of the moving object.

[0073] The current trajectory transition features corresponding to each edge in the temporal dependency graph are input into the trained Isolation Forest algorithm model. For each transition edge, its displacement, duration, and region type feature values ​​are extracted to form a feature vector input to the model. The feature vector needs to be standardized to ensure that each feature dimension has the same dimensions. The standardization formula is: feature value minus mean and then divided by standard deviation. The Isolation Forest algorithm model obtains anomaly scores by calculating the path length of this feature vector in each isolation tree. The path length represents the number of random partitions required to isolate the feature vector; anomalous samples usually have shorter path lengths. The final anomaly score is calculated by averaging the path lengths of all isolation trees, with scores ranging from 0 to 1. The closer the score is to 1, the higher the degree of anomalousness.

[0074] An anomaly score is calculated for the current trajectory transition features using the Isolation Forest algorithm, generating an individual pattern deviation for each edge in the time-series dependency graph. The individual pattern deviation is directly represented by the anomaly score; a higher score indicates that the movement pattern represented by that edge deviates more from the historical normal pattern of the moving object. For example, transition edges with an individual pattern deviation greater than 0.6 are considered anomalous transitions that significantly deviate from the normal pattern. The calculated individual pattern deviation is added as a new attribute to the transition edge in the time-series dependency graph, providing a data foundation for subsequent analysis. The entire calculation process is implemented using Python programming, utilizing the IsolationForest class from the scikit-learn machine learning library for model training and prediction. The contamination parameter is set to 0.1 to adapt to most trajectory data analysis scenarios, ensuring the accuracy and reliability of the algorithm. After model training, it is saved as a file for easy loading and use later, avoiding the computational overhead of repeated training.

[0075] S3. By analyzing the movement intention of trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold. Figure 1 Consistency, inferring the continuity of the intention of the moving object, specifically includes:

[0076] In the process of deducing the intent coherence of moving objects, the first step is to identify edges in the temporal dependency graph where the individual pattern deviation exceeds a preset threshold. The preset threshold is determined by analyzing the distribution characteristics of individual pattern deviation in the historical trajectory data of the moving object; for example, the 95th percentile of historical deviation values ​​is used as the threshold, typically ranging from 0.6 to 0.8. Statistical analysis methods are employed to determine the threshold, calculating the percentile of the historical individual pattern deviation data and selecting appropriate quantiles as the threshold, such as the 90th, 95th, or 99th percentile. The specific quantile selection depends on the sensitivity requirements for anomaly detection. The identification process traverses all transition edges in the temporal dependency graph, comparing the individual pattern deviation value of each edge with the preset threshold. Edges with individual pattern deviations exceeding the preset threshold are marked as edges to be analyzed, and their identifiers and connections are recorded, along with the specific numerical value of the individual pattern deviation for each edge.

[0077] The system extracts trajectory segments connected by edges whose individual pattern deviation exceeds a preset threshold. Each trajectory segment contains the original trajectory point sequence corresponding to the edge to be analyzed and several adjacent edges before and after it. The segment length is determined based on the activity characteristics of the moving object, for example, containing 5 to 10 consecutive trajectory points, with a corresponding duration ranging from approximately 30 seconds to 5 minutes. During the extraction process, the complete trajectory segment is reconstructed according to the timestamp order of the trajectory points to ensure temporal continuity. The extracted trajectory segments undergo data preprocessing, including removing outliers with large GPS positioning errors, such as points with horizontal accuracy greater than 10 meters; interpolating the trajectory points to ensure uniform time intervals, for example, one trajectory point every 2 seconds; and using linear interpolation to supplement the coordinates of missing trajectory points.

[0078] The movement pattern features of moving objects in the extracted trajectory segments are analyzed. These features include velocity change sequences and direction change sequences. The velocity change sequence is obtained by calculating the ratio of displacement to time interval between adjacent trajectory points, with velocity measured in meters per second (m / s). The calculation uses the Euclidean distance between the two points divided by the time difference. The direction change sequence is obtained by calculating the change in the direction angle of the line connecting adjacent trajectory points, with the direction angle measured in degrees, starting from 0 degrees north and increasing clockwise. The arctangent function is used for angle calculation. A sliding window method is employed during feature extraction, with a window size of, for example, three trajectory points and a step size of one point. Feature values ​​within each window are calculated to form a complete feature sequence. The sequence length is related to the number of trajectory points.

[0079] Calculate the statistical consistency index of the velocity change sequence of the trajectory segment. The statistical consistency index includes the variance and coefficient of variation of the velocity sequence. Variance reflects the absolute amplitude of velocity fluctuations, and the coefficient of variation reflects the relative amplitude of velocity fluctuations. The variance is calculated as the average of the sum of squares of the differences between each velocity value and the average velocity; the coefficient of variation is calculated as the ratio of the standard deviation to the mean, where the standard deviation is the square root of the variance. Simultaneously, calculate the autocorrelation coefficient of the velocity sequence to assess the regularity of velocity changes. The autocorrelation coefficient is calculated using time series analysis methods, with lag orders set to 1 to 3. The autocorrelation coefficients for each lag order are calculated, and the average is taken as the final index.

[0080] The statistical consistency index of the directional change sequence of trajectory segments is calculated. The directional consistency index includes the root mean square value of the rate of change of directional angles and the directional stability index. The rate of change of directional angles is obtained by calculating the difference between adjacent directional angles. The root mean square value reflects the average magnitude of the directional change, and is calculated as the square root of the average of the sum of squares of the rates of change. The directional stability index is obtained by calculating the proportion of time the directional angle remains within a predetermined range, for example, the proportion of time the directional change is less than 10 degrees / second. The predetermined range is determined based on the type of moving object. Simultaneously, the entropy value of the directional sequence is calculated to assess the randomness of directional changes. The entropy value is calculated as the information entropy calculated by the probability distribution function of each directional change value. The probability distribution is obtained by statistically analyzing the frequency of occurrence of each directional change value.

[0081] Based on the statistical consistency index of velocity change sequence and direction change sequence, the movement intention of trajectory segments is comprehensively analyzed. Figure 1 Consistency. A comprehensive evaluation function is established, which normalizes each statistical consistency indicator and then performs a weighted combination. The weights are determined based on the importance of the indicators; for example, velocity variance has a weight of 0.4, directional change rate has a weight of 0.3, and entropy value has a weight of 0.3, with the sum of the weights being 1. Through a pre-defined intention... Figure 1 Consistency threshold determines the movement intention of trajectory segments Figure 1 The consistency score threshold is obtained through training on historical data. For example, the 10th percentile of the consistency score of normal trajectory segments is taken as the threshold. The consistency score is calculated using a weighted summation formula. Trajectory segments with a consistency score higher than the threshold are judged to have moving intent. Figure 1 In terms of consistency, trajectory segments with scores below a threshold are judged as having inconsistent movement intentions.

[0082] According to the intention of movement Figure 1 The consistency analysis results are used to infer the intentional coherence of the moving object. Intentional coherence is determined by the movement intentions of multiple consecutive trajectory segments. Figure 1 The degree of consistency is assessed by calculating the proportion of consistent trajectory segments within a time window. The time window size is set to, for example, 10 minutes, but adjusted according to the specific application scenario. Simultaneously, the meaning is analyzed. Figure 1Consistent duration patterns are used to determine intent coherence. For example, consistency among three or more consecutive trajectory segments is considered high, while consistency among two or fewer consecutive trajectory segments is considered low. A final intent coherence score for the moving object is generated, ranging from 0 to 1. The score is calculated by multiplying the percentage of consistent trajectory segments by a duration pattern coefficient, which is determined based on the number of consecutive consistent segments. The score results are stored in association with trajectory segment information, providing input data for subsequent semantic consistency analysis. The storage format includes trajectory segment identifiers, score values, and timestamp information.

[0083] S4. By analyzing the functional attribute similarity of the regions where the nodes at both ends of the edge are located and combining it with intent coherence, a semantic consistency verification result including semantically conflicting edges is generated. The specific implementation includes:

[0084] In generating semantic consistency verification results, the functional attribute information of the regions where the nodes at both ends of the edge in the temporal dependency graph are located is first obtained. This functional attribute information is acquired through a geographic information system (GIS) and includes, but is not limited to, specific attribute data such as land use type, building function classification, and point of interest (POI) distribution characteristics. The region where each key location node is located is defined as a circular area centered on the node with a radius of, for example, 500 meters. This radius is determined based on the urban spatial structure and the activity range of moving objects; for example, a smaller radius, such as 300 meters, can be used in the city center, while a larger radius, such as 800 meters, can be used in the suburbs. The functional attribute information specifically includes dimensions such as commercial facility density, residential area ratio, industrial land ratio, and public service facility distribution. Each dimension is obtained by statistically analyzing the number or area ratio of corresponding types of POIs within the region. For example, commercial facility density is obtained by calculating the number of commercial POIs per square kilometer. The data source is publicly available geographic information data from urban planning departments or application programming interfaces (APIs) of commercial map service providers. The data acquisition frequency is determined based on regional development changes, for example, updated quarterly.

[0085] This analysis examines the similarity between the functional attribute information of the regions containing the nodes at both ends of an edge. The similarity calculation employs a cosine similarity method using multi-dimensional feature vectors. Each region's functional attribute information is represented as a feature vector, with each dimension corresponding to a different functional attribute index. For example, a vector can contain 10 dimensions, each representing the intensity value of a different functional attribute. Before similarity calculation, each functional attribute index is normalized to eliminate dimensional differences. The normalization method uses minimum-maximum scaling to convert each index value to a range of 0 to 1. The minimum and maximum values ​​are obtained based on historical data statistics. The cosine similarity calculation formula is the dot product of two feature vectors divided by the product of their respective magnitudes. The result ranges from 0 to 1; a larger value indicates greater similarity in functional attributes. A similarity threshold is set, for example, 0.6. This threshold is determined by analyzing the distribution of functional attribute similarity of normal transition edges in historical trajectory data. For example, the first quartile of the similarity value can be used as the threshold; values ​​above this threshold are considered similar in functional attributes, while values ​​below are considered dissimilar.

[0086] Based on the inferred intent coherence of mobile objects, the consistency between functional attribute similarity and intent coherence is analyzed. A consistency analysis matrix is ​​established, combining functional attribute similarity and intent coherence for analysis. Intent coherence uses the scoring results obtained in previous steps, with a score range of 0 to 1. The consistency analysis employs a four-quadrant analysis method: when functional attribute similarity is above a threshold and intent coherence score is above 0.7, it is considered highly consistent; when functional attribute similarity is below the threshold but intent coherence score is above 0.7, it is considered semantically conflicting; when functional attribute similarity is above the threshold but intent coherence score is below 0.3, it is considered potentially abnormal behavior; when functional attribute similarity is below the threshold and intent coherence score is below 0.3, it is considered low consistent but without conflict. Each threshold parameter is optimized and determined using machine learning methods, such as using grid search to find the optimal threshold combination on historical data.

[0087] Based on the consistency analysis results of functional attribute similarity and intent coherence, semantic consistency verification results, including semantically conflicting edges, are generated. Semantically conflicting edges are defined as edges where the functional attribute similarity score is below a threshold but the intent coherence score is above a threshold. These edges indicate that the moving object maintains a highly coherent movement intent between two regions with significantly different functional attributes, but semantic inconsistencies may exist. The verification results include a semantic conflict identifier, conflict severity score, and conflict type classification for each edge. The conflict severity score is calculated by the difference between the functional attribute similarity score and the intent coherence score, using a weighted difference formula. The weight coefficients are adjusted according to the moving object type; for example, pedestrians have a weight of 0.5, and vehicles have a weight of 0.7. The final semantic consistency verification results are stored in a structured data format, including edge identifiers, functional attribute similarity, intent coherence score, consistency judgment results, and semantic conflict markers, providing input for subsequent temporal dependency graph structure optimization. Storage uses JSON format and includes complete metadata information to ensure data traceability and reusability.

[0088] S5. By analyzing the multi-path substitutability decisions of semantically conflicting edges in the temporal dependency graph, the retention, weight reduction, or deletion operations of semantically conflicting edges are performed to generate a structurally optimized temporal dependency graph. Specific implementation includes:

[0089] In generating the optimized temporal dependency graph, the semantically conflicting edges in the semantic consistency verification results are first obtained. These edges are read from the semantic consistency verification result file, which is stored in a data structure containing edge identifiers, semantic conflict markers, and conflict severity scores. For example, it can be stored in JSON format and includes an `edges` field recording detailed information for all edges. Each edge's information includes a unique identifier, source node ID, target node ID, functional attribute similarity value, intent coherence score, and semantic conflict marker. During the reading process, all edges marked as semantically conflicting are filtered out based on the semantic conflict markers. Simultaneously, the source and target node identifiers of these edges, as well as the conflict severity score, are obtained. The conflict severity score is calculated using the weighted difference between the functional attribute similarity and the intent coherence score. The weighting coefficient is adjusted according to the specific application scenario; for example, a weighting coefficient of 0.6 can be used in urban traffic analysis.

[0090] For each semantically conflicting edge, a set of alternative paths connecting the two ends of the corresponding semantically conflicting edge is searched in the temporal dependency graph. The search process employs a breadth-first search algorithm, starting from the source node of the semantically conflicting edge and targeting the target node of the same edge. The search depth is limited to, for example, 3 to 5 hops, determined based on the average node distance in the urban road network to avoid excessively long search paths that would lead to high computational complexity. The set of alternative paths includes all paths that do not pass through the current semantically conflicting edge but connect the same two ends. Each path consists of a series of consecutive edges and nodes, and the path length is defined as the number of edges traversed. Paths containing the current semantically conflicting edge are excluded during the search, and a maximum limit is set on the number of paths, for example, a maximum of 10 optimal alternative paths are retained. The path selection criteria are based on a comprehensive evaluation of path length and weight to prevent the path set from becoming too large and affecting subsequent processing efficiency.

[0091] This study analyzes the reliability of paths within the set of alternative paths to assess the substitutability of semantically conflicting edges. Path reliability is calculated by multiplying the weights of all edges in the path. The weights are derived from the transition probabilities or frequency information of edges in the temporal dependency graph, with weight values ​​ranging from 0 to 1; higher weights indicate higher reliability. The assessment of multi-path substitutability considers the number of alternative paths and average reliability, setting a reliability threshold, for example, 0.7. This threshold is determined by analyzing the reliability distribution of normal paths in historical trajectory data, for example, by using the median reliability value as the threshold. Paths exceeding this threshold are considered high-reliability paths. The assessment results include indicators such as the number of alternative paths, the number of high-reliability paths, the average path reliability, and the maximum reliability. These indicators collectively determine the substitutability of semantically conflicting edges; for example, when the number of high-reliability paths is greater than 3, the multi-path substitutability is considered high.

[0092] Based on the evaluation results of the substitutability of multiple paths, decisions are made to retain, reduce the weight of, or delete semantically conflicting edges. If a semantically conflicting edge has multiple highly reliable alternative paths, such as more than three alternative paths with a reliability higher than 0.7, the semantically conflicting edge is deleted because there are enough alternative paths to maintain the connectivity of the graph. If a semantically conflicting edge has no alternative paths, the edge is retained but its weight is reduced. The reduction in weight is determined based on the conflict severity score; for example, if the conflict severity score is above 0.7, the weight is reduced to 0.3 times the original value, and if the conflict severity score is between 0.4 and 0.7, the weight is reduced to 0.6 times the original value. If a semantically conflicting edge has a few low-reliability alternative paths, such as one or two alternative paths with a reliability lower than 0.7, the edge is reduced in weight based on the comprehensive evaluation results of the reliability of the alternative paths. The reduction in weight takes into account both the reliability and the number of alternative paths; for example, if there is one alternative path with a reliability of 0.6, the weight of the original edge is reduced to 0.5 times the original value.

[0093] Based on the results of operations on semantically conflicting edges, a structurally optimized temporal dependency graph is generated. For deleted semantically conflicting edges, they are removed from the edge set of the temporal dependency graph, and the adjacency list data structure is updated to ensure that the graph's connectivity is unaffected. For semantically conflicting edges that have been downweighted, their weight values ​​in the temporal dependency graph are modified, and the weights are re-normalized to ensure that the sum of the weights of all edges originating from the same node is 1. The normalization process uses the weight divided by the sum of weights. The structurally optimized temporal dependency graph retains the original node set, while the edge set is adjusted accordingly based on the operation results. The final generated temporal dependency graph is stored as an independent graph structure file, containing complete node and edge information, as well as updated weight data, providing input for subsequent trajectory archive vector generation. The entire optimization process ensures that the temporal dependency graph maintains connectivity while eliminating the noise impact of semantically conflicting edges, improving the accuracy of trajectory representation. The optimized temporal dependency graph uses the same data structure and storage format as the original temporal dependency graph to ensure compatibility with subsequent processing steps.

[0094] S6. By analyzing the processing results of the weighted graph neural network on the optimized temporal dependency graph, a trajectory file vector of the moving object is generated. Specific implementation includes:

[0095] In generating trajectory archive vectors for moving objects, the optimized temporal dependency graph is first input into a weighted graph neural network. The input to the weighted graph neural network includes a node feature matrix and an edge weight matrix. The node feature matrix contains the feature vectors of each key node, with features including, but not limited to, node access frequency, average dwell time, and time distribution characteristics. Node access frequency is obtained by counting the number of trajectory points falling into that node in historical trajectory data. The average dwell time is obtained by calculating the average of the time differences between trajectory points that stay at that node. The time distribution characteristics are obtained by analyzing the time distribution of visits to that node, for example, using the access frequency distribution over a 24-hour period. The edge weight matrix is ​​derived from the weight values ​​of each edge in the optimized temporal dependency graph, and these weight values ​​have been updated after previous weight reduction and deletion operations. The weighted graph neural network adopts a multi-layer graph convolutional structure, with the number of layers set to, for example, 2 to 4 layers. The specific number of layers is determined according to the number of nodes and graph density of the temporal dependency graph. Each layer contains a linear transformation and a non-linear activation function. The activation function is the ReLU function. The neural network parameters are obtained through backpropagation. The training objective is to minimize the trajectory reconstruction error. A learning rate decay strategy is used during training. The initial learning rate is set to, for example, 0.01. Every 50 training rounds, the learning rate decays to half of its original value.

[0096] The neighborhood feature information of nodes in the optimized temporal dependency graph is aggregated using a weighted graph neural network. The weights of neighboring nodes are calculated based on the edge weights in the optimized graph. The calculation process first normalizes the outgoing edge weights of each node, ensuring the sum of all outgoing edge weights is 1. Normalization is achieved by dividing the weight of each edge by the sum of all outgoing edge weights of that node, ensuring the weight distribution conforms to the probability distribution requirements. The feature vectors of each neighboring node of the target node are then weighted and summed according to their respective neighboring node weights. The weighted summation formula is to multiply each neighboring node's feature vector by its corresponding normalized weight and then sum them, maintaining consistency in the feature dimensions during the summation process. The weighted summation result is then linearly combined with the target node's own feature vector and subjected to nonlinear transformation. The linear combination uses a learnable weight matrix to map the two vectors to the same feature space. The dimension of the weight matrix is ​​determined based on the input and output feature dimensions; for example, if the input feature dimension is 32, the weight matrix dimension is 32×64. The nonlinear transformation uses an activation function, specifically the ReLU function, to generate a new feature vector after aggregation of the target node. The dimension of this new feature vector is consistent with the output dimension of the weight matrix.

[0097] Based on the aggregation results of node neighborhood feature information by a weighted graph neural network, node representation vectors are generated. The generation of node representation vectors is completed iteratively through a multi-layer graph neural network, with the output of each layer serving as the input to the next layer, and the output of the final layer becoming the node representation vector. The dimension of the node representation vector is set to, for example, 64 or 128 dimensions, determined by the size of the temporal dependency graph; a larger dimension is used when the number of nodes in the graph is large, for example, a 128-dimensional feature vector is used when the number of nodes exceeds 1000. A skip connection mechanism is employed during generation to fuse initial node features with intermediate layer features, ensuring effective preservation of features at different granularities. Skip connections are implemented through feature concatenation, for example, concatenating the output features of layers 1 and 2 along the feature dimension. The node representation vector contains structural and semantic information of the node, representing the importance and functional characteristics of the node in the temporal dependency graph. This information is fully extracted and compressed through multi-layer transformation and aggregation of the neural network.

[0098] The aggregation of node representation vectors from all nodes in the time-dependent graph generates a trajectory profile vector for the moving object. The aggregation operation employs a weighted summation method based on an attention mechanism. First, the attention weight of each node is calculated using a neural network. This network takes the node representation vector as input and outputs a scalar weight value. This weight value is then normalized using a softmax function, ensuring the sum of the attention weights for all nodes is 1. The weighted summation formula is the sum of the sums of the node representation vectors multiplied by their corresponding attention weights, resulting in a fixed-dimensional trajectory profile vector. The dimension of the trajectory profile vector is consistent with the dimension of the node representation vectors, such as 64 or 128 dimensions. This vector comprehensively represents the movement patterns and behavioral characteristics of the moving object. The final generated trajectory profile vector is normalized using L2 normalization to ensure that the feature values ​​of each dimension are within the same numerical range, facilitating subsequent similarity calculations and pattern recognition applications. The trajectory profile vector is stored in a feature file format, such as a NumPy array, and includes a generation timestamp and version information to ensure the traceability and reusability of the trajectory profile. The entire generation process is implemented using Python programming. A weighted graph neural network is built using the PyTorch deep learning framework, and the Adam optimizer is used for model training. The number of training rounds is set to 200, and the batch size is set to 32 or 64 depending on the memory capacity.

[0099] Example 2: Figure 2 A schematic diagram of the trajectory file construction system based on temporal dependency graphs and neural networks of the present invention is given. The trajectory file construction system based on temporal dependency graphs and neural networks includes:

[0100] The proximity analysis module is used to generate a temporal dependency graph containing nodes and transition edges by analyzing the spatiotemporal proximity relationships in the original trajectory data of moving objects.

[0101] The deviation generation module is used to analyze the degree of conformity between the current trajectory transfer characteristics and the historical trajectory pattern characteristics of the moving object by using the isolated forest algorithm, and generate the individual pattern deviation of each edge in the time-series dependency graph.

[0102] The coherent deduction module is used to analyze the movement intention of trajectory segments connected by edges whose deviation from individual patterns exceeds a preset threshold. Figure 1 Consistency, inferring the continuity of the intention of the moving object;

[0103] The semantic verification module is used to generate semantic consistency verification results, including semantically conflicting edges, by analyzing the functional attribute similarity of the regions where the nodes at both ends of the edge are located and combining it with intent coherence.

[0104] The optimization processing module is used to generate a structurally optimized temporal dependency graph by analyzing the multi-path substitutability decisions of semantically conflicting edges in the temporal dependency graph and performing operations such as retention, weight reduction or deletion of semantically conflicting edges.

[0105] The archive generation module is used to generate trajectory archive vectors of moving objects by analyzing the processing results of the time-series dependency graph after structural optimization by a weighted graph neural network.

[0106] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0107] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0108] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0109] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0111] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0112] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0113] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0115] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A trajectory profile construction method based on a timing dependency graph and a neural network, characterized in that, The method comprises the following steps: S1, generating a time sequence dependency graph containing nodes and transition edges by analyzing the spatio-temporal proximity relationship in the original trajectory data of the moving object; S2, generating the individual mode deviation degree of each edge in the time sequence dependency graph by analyzing the degree of conformity between the current trajectory transfer feature and the historical trajectory mode feature of the moving object using the isolation forest algorithm, including: extracting the current trajectory transfer feature corresponding to the transition edge in the time sequence dependency graph, the current trajectory transfer feature including displacement, time length and region type; obtaining the trajectory data of the moving object in the historical period, and constructing a historical trajectory mode feature set of the moving object according to the historical trajectory data; inputting the historical trajectory mode feature set of the moving object into the isolation forest algorithm for model training; inputting the current trajectory transfer feature corresponding to each edge in the time sequence dependency graph into the trained isolation forest algorithm model; calculating the anomaly score of the current trajectory transfer feature through the isolation forest algorithm model to generate the individual mode deviation degree of each edge in the time sequence dependency graph; S3, deducing the intention continuity of the moving object by analyzing the moving intention consistency of the trajectory segments connected by the edges with the individual mode deviation degree higher than the preset threshold, including: identifying the edges in the time sequence dependency graph with the individual mode deviation degree higher than the preset threshold; extracting the trajectory segments connected by the edges with the individual mode deviation degree higher than the preset threshold; analyzing the moving mode features of the extracted trajectory segments, including speed change sequence and direction change sequence; analyzing the moving intention consistency of the trajectory segments according to the moving mode features, including: calculating the statistical consistency index of the speed change sequence of the trajectory segment, calculating the statistical consistency index of the direction change sequence of the trajectory segment, establishing a comprehensive evaluation function, weighting and combining the statistical consistency indexes of the speed change sequence and the direction change sequence after normalization to obtain the moving intention consistency score of the trajectory segment, and judging the moving intention consistency degree of the trajectory segment through the moving intention consistency score and the preset intention consistency threshold; deducing the intention continuity of the moving object according to the analysis result of the moving intention consistency, including: evaluating the intention continuity score of the moving object through the moving intention consistency degrees of the continuous multiple trajectory segments; S4, generating the semantic consistency verification result including the semantic conflict edges by analyzing the functional attribute similarity of the regions where the nodes at both ends of the edges are located and combining the intention continuity, including: obtaining the functional attribute information of the regions where the nodes at both ends of the edges in the time sequence dependency graph are located; analyzing the similarity between the functional attribute information of the regions where the nodes at both ends of the edges are located; combining the deduced intention continuity of the moving object to analyze the consistency of the functional attribute similarity and the intention continuity; generating the semantic consistency verification result including the semantic conflict edges according to the analysis result of the consistency of the functional attribute similarity and the intention continuity, wherein the semantic conflict edges are the edges with the functional attribute similarity lower than the threshold but the intention continuity score higher than the threshold; S5, generating the time sequence dependency graph after structure optimization by analyzing the reservation, weight reduction or deletion operation of the semantic conflict edges in the time sequence dependency graph through the multi-path replaceability decision of the semantic conflict edges. S6, generating a trajectory profile vector of the moving object by analyzing a processing result of the structure-optimized time-dependent graph by the weighted graph neural network, comprising: inputting the structure-optimized time-dependent graph into the weighted graph neural network; aggregating neighborhood feature information of nodes in the structure-optimized time-dependent graph by the weighted graph neural network; generating a node representation vector according to the aggregation result of the neighborhood feature information of the nodes by the weighted graph neural network; aggregating the node representation vectors of all nodes in the time-dependent graph to generate the trajectory profile vector of the moving object. 2.The trajectory profile construction method based on time-dependent graph and neural network according to claim 1, wherein, S1, generating a time-dependent graph containing nodes and transfer edges by analyzing the spatio-temporal proximity relationship in the original trajectory data of the moving object, comprising: performing trajectory point clustering on the original trajectory data of the moving object to generate key position nodes by clustering trajectory points that are adjacent in space; analyzing the transfer relationship between the key position nodes to which adjacent trajectory points belong according to the time sequence of the trajectory points to generate transfer edges between the key position nodes; constructing the time-dependent graph by the key position nodes and the transfer edges. 3.The trajectory profile construction method based on time-dependent graph and neural network according to claim 1, wherein, S3, generating a structure-optimized time-dependent graph by analyzing the multi-path replaceability decision of the semantic conflict edge in the time-dependent graph to retain, weight down or delete the semantic conflict edge, comprising: obtaining the semantic conflict edge contained in the semantic consistency verification result; for each semantic conflict edge, searching for a set of alternative paths connecting the two end nodes of the corresponding semantic conflict edge in the time-dependent graph; analyzing the reliability of the paths in the set of alternative paths to evaluate the multi-path replaceability degree of the semantic conflict edge; according to the evaluation result of the multi-path replaceability degree, deciding to perform the retention, weight down or deletion operation on the semantic conflict edge: if the semantic conflict edge has multiple high-reliability alternative paths, the semantic conflict edge is deleted; if the semantic conflict edge has no alternative path, the semantic conflict edge is retained but the weight of the semantic conflict edge is reduced; if the semantic conflict edge has a few low-reliability alternative paths, the weight of the semantic conflict edge is reduced according to the comprehensive evaluation result of the reliability of the alternative paths; generating a structure-optimized time-dependent graph based on the operation result of the semantic conflict edge. 4.The trajectory profile construction method based on time-dependent graph and neural network according to claim 1, wherein, S4, aggregating the neighborhood feature information of the nodes in the structure-optimized time-dependent graph by the weighted graph neural network, comprising: calculating the neighbor node weight according to the weight of the edge in the structure-optimized time-dependent graph; weighting and summing the feature vectors of each neighbor node of the target node according to the corresponding neighbor node weight; linearly combining and nonlinearly transforming the weighting and summing result with the feature vector of the target node to generate a new feature vector of the target node after aggregation.

5. A trajectory profile construction system based on time-dependent graph and neural network for implementing the trajectory profile construction method based on time-dependent graph and neural network according to any one of claims 1-4, characterized in that, comprising: a proximity analysis module for generating a time-dependent graph containing nodes and transfer edges by analyzing the spatio-temporal proximity relationship in the original trajectory data of the moving object; a deviation generation module for generating an individual mode deviation degree of each edge in the time-dependent graph by analyzing the degree of conformity of the current trajectory transfer feature with the historical trajectory mode feature of the moving object using the isolation forest algorithm; a coherent inference module for inferring the intention coherence of the moving object by analyzing the consistency of the moving intention of the trajectory segment connected by the edge with the individual mode deviation degree higher than the preset threshold. The semantic verification module is configured to generate a semantic consistency verification result including the semantic conflict edges by analyzing the similarity of the functional attributes of the areas where the nodes at the two ends of the edges are located and combining the intention coherence; The optimization processing module is configured to generate a structure-optimized temporal dependency graph by analyzing the multi-path replaceability decision of the semantic conflict edges in the temporal dependency graph and performing a reservation, weight reduction or deletion operation on the semantic conflict edges; The archive generation module is configured to generate a trajectory archive vector of the mobile object by analyzing the processing result of the structure-optimized temporal dependency graph by the weighted graph neural network.

Images