Illegal operating vehicle identification system and method based on space-time diagram neural network

Abstract

The invention relates to the technical field of intelligent traffic, in particular to an illegal operating vehicle recognition system and method based on a space-time diagram neural network, and the system comprises a data interaction module, a preprocessing module, a space-time diagram construction module, a topology enhancement hierarchical space-time attention network module, a traffic feature extraction module, a data fusion module and an intelligent decision module. The data interaction module is connected with multiple platforms to obtain various data, the preprocessing module unifies the format and codes the data in a standardized mode, the space-time diagram construction module constructs a space-time dynamic relation graph containing multiple nodes, and the topology enhancement hierarchical space-time attention network extracts space-time joint feature vectors. The traffic feature extraction module extracts travel frequency and travel coherence features based on checkpoint data and calculates illegal operation probability, the data fusion module fuses the features to extract candidate illegal operating vehicles to form a vehicle library, the intelligent decision module calculates a feature weight coefficient, and the illegal operating vehicles are screened and determined based on an entropy method. And the identification coverage rate of illegal operating vehicles is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation, and particularly to an illegal operation vehicle identification system and method based on a spatio-temporal graph neural network. Background Art

[0002] With the development of the sharing economy, the number of illegal operation vehicles has been increasing day by day. These vehicles do not have legitimate operation qualifications, which not only disrupt the normal taxi market order but also pose potential safety hazards to passengers. Traditional methods for identifying illegal operation vehicles mainly rely on manual inspections, with low efficiency and limited coverage.

[0003] Existing technical solutions mainly include vehicle identification based on bayonet cameras, behavior analysis based on GPS trajectories, and post-event processing based on passenger reports, etc. These methods usually have the following problems: First, the information from a single data source is limited, making it difficult to comprehensively capture the behavioral characteristics of illegal operation vehicles; second, traditional feature extraction methods are difficult to effectively process the spatio-temporal correlation of vehicle behaviors, resulting in low identification accuracy; third, there is a lack of effective evidence storage and management mechanisms, making it difficult to obtain law enforcement evidence.

[0004] In addition, existing methods usually adopt simple data fusion strategies and cannot make full use of the complementary characteristics of multi-source heterogeneous data; during the feature extraction process, ordinary neural networks are difficult to capture complex spatio-temporal structure information; in the decision-making stage, a single threshold is difficult to adapt to the differentiated requirements of different cities and scenarios. These problems seriously limit the practical application effect of the illegal operation vehicle identification system. Summary of the Invention

[0005] The purpose of the present invention is to provide an illegal operation vehicle identification system and method based on a spatio-temporal graph neural network. By introducing a topology-enhanced hierarchical spatio-temporal attention network, high-precision identification of illegal operation vehicles is achieved, effectively solving problems such as incomplete feature extraction, low identification accuracy, and imperfect evidence management in the prior art.

[0006] The present invention proposes an illegal operation vehicle identification system based on a spatio-temporal graph neural network, including: A data interaction module, used to connect to a mobile payment management platform, an urban traffic control platform, and a blockchain evidence storage platform, and obtain bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data; A data preprocessing module, connected to the data interaction module, used to unify the formats of the bayonet capture data, the mobile payment platform trip data, the urban traffic control platform data, and the passenger APP operation behavior data, and perform encoding standardization on each dimension of data; A spatio-temporal graph construction module, connected to the data preprocessing module, is used to construct a spatio-temporal dynamic relationship graph including driver nodes, checkpoint nodes, vehicle nodes, vehicle type nodes, trip nodes and parking point nodes based on the encoded and standardized data; A topological enhanced hierarchical spatio-temporal attention network, connected to the spatio-temporal graph construction module, includes a topological space mapping layer, a probability attention weighting layer and a spatio-temporal feature fusion layer, and is used to extract a spatio-temporal joint feature vector from the spatio-temporal dynamic relationship graph; A traffic feature extraction module, connected to the data preprocessing module, is used to extract travel frequency and trip coherence features based on the time-space trajectory of the checkpoint capture data, and calculate the probability of illegal operation; A data fusion module, connected to the topological enhanced hierarchical spatio-temporal attention network and the traffic feature extraction module, is used to fuse the spatio-temporal joint feature vector, the travel frequency and the trip coherence feature, extract candidate illegal operation vehicles, and form an illegal operation vehicle library; An intelligent decision-making module, connected to the data fusion module, is used to calculate the weight coefficients of each dimension feature, screen the vehicles in the illegal operation vehicle library based on the entropy value method, and determine the illegal operation vehicles.

[0007] Preferably, the topological space mapping layer is used to: Represent the urban road network as a graph structure, where intersections are nodes and roads are edges; Assign weights to each edge of the graph structure, and the weights represent road travel time or travel difficulty; Construct a road adjacency matrix to represent the connectivity of the road network; Map the vehicle GPS trajectory point sequence to the nearest road edge; Calculate the path representation of the trajectory in the topological space; Extract the topological invariant features of the path; Calculate the topological persistence index of the trajectory, representing the trajectory structure stability; Identify the topological features in the trajectory; Generate a trajectory topological signature.

[0008] Preferably, the probability attention weighting layer is used to: Decompose the trajectory into sub-trajectories of different time scales; Extract features for each time scale to capture behavior patterns of different time granularities; Construct a multi-scale feature matrix to represent the feature distribution of different time scales; Construct a conditional probability model to quantify the importance of features for illegal operation judgment; Calculate the information gain of each feature as the basis for the attention weight; Design three attention heads: time, space, and spatiotemporal joint attention heads; The attention mechanism is used to highlight abnormal behavior characteristics and improve detection sensitivity.

[0009] Preferably, the spatiotemporal feature fusion layer is used for: Analyze the complementarity and redundancy of characteristics from different sources; Remove highly correlated features and retain complementary information; Construct feature dependency graph to guide feature fusion strategy; Adjust the fusion weights based on city characteristics, time period, weather and other factors; Construct joint features representing the starting area, travel path, passing areas, etc.; Preserve high-order correlations between features; Generate the final spatiotemporal joint feature vector.

[0010] Preferably, the data preprocessing module is further used for: Standardize time information and unify the time representation format of data from different sources; Standardize spatial information and map location information to a unified geographic coordinate system; Standardize vehicle information and unify vehicle identification representations from different sources; Perform outlier detection and processing to identify and process data that significantly deviates from the normal range; Perform missing value completion and make reasonable inferences about missing data based on spatiotemporal correlation; Perform data consistency verification to check whether the same information from different sources is consistent.

[0011] Preferably, the space-time graph construction module is further used for: Set driving habits, driving time and other attributes for the driver node; Set geographic location, passing frequency and other attributes for the checkpoint node; Set attributes such as vehicle model and license plate color for vehicle nodes; Set attributes such as function type and regional characteristics for location nodes; Construct vehicle-checkpoint edges to represent the time and frequency of vehicles passing through the checkpoints; Construct vehicle-location edges to represent the vehicle’s stay time and frequency at the location; Construct driver-vehicle edges to represent driving association strength and time distribution; Construct location-location edges to represent the travel frequency and travel time between locations; Implement dynamic graph updates and update the graph structure when new data arrives.

[0012] Preferably, the traffic feature extraction module is further configured to: Calculate the travel frequency and count the number of trips within different time windows; Analyze the trip coherence and evaluate the spatio-temporal logical rationality between consecutive trips; Extract the parking pattern and identify regular parking points and duration distributions; Calculate the Gini coefficient of the trip trajectory to quantify the inequality of travel distances; Set a Gini coefficient threshold to determine the identity of professional drivers; Divide the urban area into several sectors; Use the boundary checkpoints of each sector to construct a spatio-temporal feature set; Cluster the boundary checkpoints into several clusters according to the feature similarity of the feature set; Calculate the average residence time of checkpoints within the cluster and calculate the probability of illegal operation.

[0013] Preferably, the system further includes: A passenger APP feature extraction module, connected to the data interaction module, for extracting user rating patterns, search behaviors, and payment records from the passenger APP operation behavior data and calculating the possibility of illegal operation; The data fusion module is further configured to fuse the possibility of illegal operation.

[0014] Preferably, the system further includes: A blockchain evidence storage module, connected to the intelligent decision-making module, for receiving and encrypting and storing the identification results of illegal operation vehicles, including information such as MAC addresses, usage times, license plates, and driving routes; Perform block storage according to a preset time period; Provide verifiable law enforcement evidence.

[0015] An illegal operation vehicle identification method based on a spatio-temporal graph neural network, using the described system, is characterized by including: Connect to the mobile payment management platform, urban traffic control platform, and blockchain evidence storage platform through the data interaction module to obtain checkpoint capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data; Unify the formats of the checkpoint capture data, the mobile payment platform trip data, the urban traffic control platform data, and the passenger APP operation behavior data through the data preprocessing module, and perform encoding standardization on the data of each dimension; Construct a spatio-temporal dynamic relationship graph including driver nodes, checkpoint nodes, vehicle nodes, vehicle type nodes, trip nodes, and parking point nodes based on the encoded and standardized data through the spatio-temporal graph construction module; Extract spatio-temporal joint feature vectors from the spatio-temporal dynamic relationship graph through a topology-enhanced hierarchical spatio-temporal attention network, where the topology-enhanced hierarchical spatio-temporal attention network includes a topological space mapping layer, a probability attention weighting layer, and a spatio-temporal feature fusion layer; Through the traffic feature extraction module, extract the travel frequency and trip coherence features based on the time-space trajectory of the bayonet capture data, and calculate the probability of illegal operation; Through the data fusion module, fuse the spatio-temporal joint feature vectors, the travel frequency, and the trip coherence features, extract candidate illegal operation vehicles, and form an illegal operation vehicle library; Through the intelligent decision-making module, calculate the weight coefficients of each dimension feature, and screen the vehicles in the illegal operation vehicle library based on the entropy method to determine illegal operation vehicles.

[0016] The present invention has the following beneficial effects: 1. Through the multi-source heterogeneous data fusion mechanism, integrating bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data, it realizes a more comprehensive capture of vehicle behavior characteristics, and greatly improves the recognition coverage rate of illegal operation vehicles.

[0017] 2. Adopt an innovative topology-enhanced hierarchical spatio-temporal attention network, and realize the extraction of the essential structural features of vehicle trajectories through a three-layer progressive structure (topological space mapping layer, probability attention weighting layer, and spatio-temporal feature fusion layer). Compared with traditional methods, the recognition accuracy is increased by about 30%, and the misjudgment rate is reduced by about 50%.

[0018] 3. Introduce blockchain evidence storage technology to ensure the reliability and immutability of law enforcement evidence, and greatly improve the success rate of illegal operation law enforcement.

[0019] 4. The system has an adaptive characteristic, can automatically adjust the model parameters according to different city characteristics, seasonal changes, and weather factors, and significantly improves the robustness and adaptability of the system.

[0020] 5. The multi-dimensional feature weight optimization mechanism based on the entropy method realizes the automatic evaluation and dynamic adjustment of feature importance, avoids the subjectivity of manually setting weights, and improves the objectivity and fairness of the system. Brief Description of the Drawings

[0021] Figure 1 It is the overall architecture diagram of the illegal operation vehicle recognition system based on the spatio-temporal graph neural network of the present invention; Figure 2 It is the structural schematic diagram of the topology-enhanced hierarchical spatio-temporal attention network of the present invention; Figure 3 It is the workflow diagram of the spatio-temporal graph construction module of the present invention; Figure 4 Schematic diagram of feature fusion of the data fusion module of the present invention; Figure 5 Decision flow chart of the intelligent decision-making module of the present invention; Figure 6 Flow chart of the method for identifying illegal operating vehicles of the present invention. Detailed implementation manners

[0022] Please refer to the attached Figure 1-6 drawings. Hereinafter, with reference to the drawings and specific embodiments, the present invention will be further described in detail.

[0023] As Figure 1 shown, the illegal operating vehicle identification system based on spatio-temporal graph neural network provided by the present invention includes a data interaction module 10, a data preprocessing module 20, a spatio-temporal graph construction module 30, a topological enhanced hierarchical spatio-temporal attention network 40, a traffic feature extraction module 50, a data fusion module 60, an intelligent decision-making module 70, a passenger APP feature extraction module 80, and a blockchain evidence storage module 90.

[0024] The data interaction module 10 is used to connect to a mobile payment management platform, an urban traffic control platform, and a blockchain evidence storage platform, and obtain bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data. Preferably, the data interaction module 10 uses a standard API interface to exchange data with each platform, and supports two modes of real-time data stream and batch data acquisition. In an embodiment of the present invention, the data interaction module 10 obtains trip data from the mobile payment platform once every 5 minutes, and obtains bayonet capture data from the urban traffic control platform once every 1 minute to ensure the timeliness of the data.

[0025] The data preprocessing module 20 is connected to the data interaction module 10, and is used to unify the formats of the bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data, and perform encoding standardization on the data of each dimension. Preferably, this module performs unified formatting processing on time information, using the standard time format of YYYY-MM-DD HH:MM:SS; performs coordinate system conversion on spatial information, and uniformly uses the WGS84 coordinate system; performs standardization processing on license plate information, removing special characters and spaces to facilitate subsequent matching and analysis. In addition, the data preprocessing module 20 is also responsible for identifying and processing abnormal data, such as GPS coordinates that deviate significantly from the reasonable range, records with incorrect timestamps, etc.

[0026] The spatio-temporal graph construction module 30, connected to the data preprocessing module 20, is used to construct a spatio-temporal dynamic relationship graph containing driver nodes, checkpoint nodes, vehicle nodes, vehicle type nodes, trip nodes, and parking point nodes based on the encoded and standardized data. In the spatio-temporal dynamic relationship graph, nodes represent entities, and edges represent the relationships between entities. Specifically, vehicle-checkpoint edges are constructed to represent the time and frequency of a vehicle passing through a checkpoint; vehicle-location edges are constructed to represent the residence time and frequency of a vehicle at a location; driver-vehicle edges are constructed to represent the driving association strength and time distribution; location-location edges are constructed to represent the trip frequency and passing time between locations. At the same time, the spatio-temporal graph construction module 30 also supports dynamic graph updates, and updates the graph structure in a timely manner when new data arrives to ensure the timeliness of the graph.

[0027] The topology-enhanced hierarchical spatio-temporal attention network 40, connected to the spatio-temporal graph construction module 30, includes a topological space mapping layer 41, a probability attention weighting layer 42, and a spatio-temporal feature fusion layer 43, and is used to extract spatio-temporal joint feature vectors from the spatio-temporal dynamic relationship graph. Preferably, the spatio-temporal joint feature vector is a vector with a dimension of 128, containing the spatio-temporal behavior characteristics of the vehicle, and is used for subsequent illegal operation judgment.

[0028] The traffic feature extraction module 50, connected to the data preprocessing module 20, is used to extract travel frequency and trip coherence features based on the time-space trajectory of checkpoint capture data and calculate the probability of illegal operation. Specifically, this module counts the number of trips of a vehicle within different time windows, such as the travel frequency distribution within 24 hours, on weekdays, weekends, etc.; analyzes the spatio-temporal logical rationality between consecutive trips to determine whether there are abnormal consecutive trips; identifies the regular parking points and parking duration distribution of the vehicle to determine whether the vehicle has professional operation characteristics.

[0029] The data fusion module 60, connected to the topology-enhanced hierarchical spatio-temporal attention network 40 and the traffic feature extraction module 50, is used to fuse the spatio-temporal joint feature vectors, travel frequency, and trip coherence features, extract candidate illegal operation vehicles, and form an illegal operation vehicle library. Preferably, the data fusion module 60 uses a weighted fusion method, assigns weights according to the discrimination ability of different features, generates a comprehensive score, and screens out candidate illegal operation vehicles according to the score.

[0030] The intelligent decision-making module 70, which is connected to the data fusion module 60, is used to calculate the weight coefficients of the features in each dimension, screen the vehicles in the illegal operation vehicle library based on the entropy method, and determine the illegal operation vehicles. In the embodiment of the present invention, the entropy method automatically assigns weights according to the information content of the features, avoiding the deviation of subjectively setting weights. Preferably, the intelligent decision-making module 70 sets the comprehensive scoring threshold to 0.75 (with a full score of 1), and the vehicles with a score higher than this threshold are determined to be illegal operation vehicles. The setting of this threshold is based on a large number of historical case analyses and expert experience, which can control the misjudgment rate while ensuring the recognition rate.

[0031] The topological enhanced hierarchical spatio-temporal attention network 40 of the present invention is the core innovation point of the system, including three levels: the topological space mapping layer 41, the probability attention weighting layer 42, and the spatio-temporal feature fusion layer 43, and realizes high-quality spatio-temporal feature extraction in a progressive manner.

[0032] The implementation method of the topological space mapping layer 41 is as follows: First, represent the urban road network as a graph structure , where is the set of intersection nodes, is the set of road edges. For each road , assign a weight , representing the travel time or travel difficulty of the road. In actual implementation, the weight can be calculated based on factors such as road length, travel speed, and congestion situation. For example: , where is the road length (unit: meter), is the average travel speed of the road (unit: meter / second), is the road congestion index (value range: between 0 and 1), is the congestion impact factor, used to adjust the impact of congestion on travel time, and generally takes a value of 0.5.

[0033] Next, construct the road adjacency matrix , representing the connectivity of the road network, where the element is defined as: , where represents the element in the matrix at the -th row and the -th column, and the value of 1 indicates that there is a road connection between node and node , and the value of 0 indicates no connection. The dimension of the matrix is , i.e., the square of the number of nodes.

[0034] Then, map the vehicle GPS trajectory point sequence to the nearest road edge to obtain a road segment sequence . To improve the mapping accuracy, use a hidden Markov model for trajectory matching, considering the road network topology constraints to ensure the continuity of the mapping results.

[0035] Among them, represents the th point in the GPS trajectory, including longitude, latitude, and timestamp information, is the total number of GPS points; represents the th road segment in the mapped road segment sequence, corresponding to an edge in the road network, is the length of the road segment sequence. Usually , because multiple adjacent GPS points may be mapped to the same road segment. Based on the mapped road segment sequence, calculate the topological persistence index of the trajectory, representing the stability of the trajectory structure: , Among them, is the persistence value of the th topological feature in the trajectory, quantifying the stability of this feature during the trajectory deformation process. The larger the value, the more stable it is; is the number of topological features considered, generally taking , corresponding to 0-dimensional, 1-dimensional, and 2-dimensional topological features (representing connected components, loops, and holes respectively).

[0036] By analyzing the topological structure of the trajectory, identify specific patterns, such as illegal operation features like detours and loops. Finally, generate a trajectory topological signature , as the basis for distinguishing different behavior patterns: , Among them, is the aforementioned topological persistence index, is the number of connected components of the trajectory, representing the number of unconnected parts in the trajectory, is the number of loops in the trajectory, representing the number of closed loops in the trajectory. is a three-dimensional vector. In practical applications, it can be further extended to a high-dimensional vector, including more topological features.

[0037] The implementation of the probability attention weighted layer 42 is as follows: First, decompose the trajectory into a set of sub-trajectories at different time scales , corresponding to hourly, daily, and weekly levels respectively. Extract features for each time scale to construct a multi-scale feature matrix : , Among them, represents the feature vector extracted from the trajectory of the time scale , representing the three time scales of hour, day, and week respectively; is the multi-scale feature matrix, with the dimension of , where is the dimension of a single feature vector (such as 128), and 3 represents the three time scales.

[0038] Then, construct a conditional probability model , quantifying the importance of features for the judgment of illegal operation, , Among them, is a binary variable indicating whether it is an illegal operation (1 means yes, 0 means no), is the th feature, is the weight parameter of this feature, learned from the training data; represents the probability that the vehicle is engaged in illegal operation under the condition of the given feature , and the value range is [0, 1]. Calculate the information gain of each feature , as the basis for the attention weight: , Among them, is the entropy of the target variable , representing a measure of uncertainty, and the calculation formula is is the conditional entropy of the target variable after the given feature , and the calculation formula is ; represents the information gain provided by the feature , and the larger the value, the greater the contribution of this feature to the judgment of .

[0039] Design three attention heads for time, space, and spatio-temporal joint, respectively focusing on features of different dimensions: The time attention head focuses on time regularity features, such as peak-hour activity frequency, night-time activity patterns, etc.; The space attention head focuses on spatial distribution features, such as frequently visited locations, activity ranges, etc.; The spatio-temporal joint attention head focuses on spatio-temporal interaction features, such as spatial distribution during specific time periods, etc.

[0040] Finally, a weighted feature vector is generated through the attention mechanism: , where is the attention weight of the th feature, satisfying , is the total number of features; is the th feature; is the weighted feature vector, with the same dimension as a single . The attention weight is calculated as follows: , where represents the natural exponential function, is the aforementioned information gain; this formula implements softmax normalization to ensure that the sum of all weights is 1.

[0041] The spatio-temporal feature fusion layer 43 is implemented as follows: First, analyze the complementarity and redundancy of features from different sources, and calculate the correlation matrix : , where is the covariance of features and , and the calculation formula is , represents the expectation operation, and are the means of features and respectively; and are the standard deviations of features and respectively; represents the Pearson correlation coefficient of features and , with a value range of [-1, 1], and the larger the absolute value, the stronger the correlation.

[0042] Remove highly correlated features and retain complementary information. Generally, when , consider removing features with lower information content to construct a feature dependence graph , where is the set of feature nodes, and each node corresponds to a feature is the set of edges of the feature dependence relationship, and the edge weight is the absolute value of the correlation coefficient.

[0043] where represents the absolute value of the correlation coefficient; the threshold 0.85 is set based on experience and represents the judgment criterion for high correlation. The number of nodes of is equal to the number of features, and the number of edges depends on the correlation distribution among the features. The fusion weights are adjusted according to factors such as urban characteristics, time periods, and weather. For example, for the core area of a large city, the spatial attention weight is appropriately increased; for the peak period on weekdays, the time attention weight is increased; for adverse weather conditions, the weights of the corresponding features are appropriately reduced.

[0044] Finally, a joint feature vector representing the starting area, travel route, passing areas, etc. is constructed. : , where , and are the time, space, and spatio-temporal joint feature vectors respectively, and their dimensions may be different; Concat represents the vector concatenation operation, which concatenates the three vectors end to end to form a longer vector; is the final joint feature vector, and its dimension is the sum of the dimensions of the three input vectors. To retain the high-order correlations among the features, tensor decomposition methods can be further used to extract potential feature patterns.

[0045] The detailed implementation method of the data preprocessing module 20 is as follows: Standardize the time information to unify the time representation formats of data from different sources. Specifically, convert all time data to the YYYY-MM-DD HH:MM:SS format and unify it to the UTC+8 time zone (Beijing time). For data lacking time zone information, infer the possible time zone based on the geographical location.

[0046] Standardize the space information by mapping the location information to a unified geographic coordinate system. Preferably, use the WGS84 coordinate system as the standard coordinate system, and perform coordinate conversion for data using other coordinate systems (such as GCJ-02, BD-09, etc.). The specific conversion formulas are as follows: Conversion from GCJ-02 to WGS84: , , where and are the latitude and longitude (in degrees) in the GCJ-02 coordinate system, and are the latitude and longitude (in degrees) in the WGS84 coordinate system; and respectively represent the latitude and longitude differences between two coordinate systems. The true WGS84 coordinates are approximated through an iterative method, and generally 10 iterations can achieve centimeter-level accuracy.

[0047] Standardize vehicle information to unify the representation of vehicle identifiers from different sources. The license plate number is unified into the format of uppercase letters plus numbers, removing special characters and spaces. For new energy vehicle license plates, their special identifiers are retained. Conduct outlier detection and handling to identify and process data that significantly deviates from the normal range. Use the Z-score method to identify outliers: , where, is the value to be detected, is the mean of this feature, is the standard deviation; is the value after standardization, also known as the Z-score. When (i.e., the absolute value is greater than 3), it is regarded as an outlier. This threshold is based on the empirical rule in statistics. Under the assumption of normal distribution, about 99.7% of the data points fall within 3 standard deviations of the mean range. For outliers, methods such as replacing with adjacent values, replacing with the mean, or directly removing can be used for handling.

[0048] Reasonably supplement missing data and infer missing data based on spatio-temporal correlation. Use the K-nearest neighbor (KNN) method for filling missing values: , where, is the missing value to be filled, is the corresponding feature value of the th sample that is most similar to the missing value sample, is the number of neighboring samples considered, usually taking 3 - 5. The similarity is calculated based on known features, and usually the Euclidean distance or Manhattan distance is used for measurement. Conduct data consistency verification to check whether the same information from different sources is consistent. For example, compare the time and location information in the bayonet capture data and the itinerary data. When the degree of inconsistency exceeds the threshold, use the data source with higher credibility. Specifically, define the consistency metric : , where, and are the same information from two sources (such as location, time, etc.), is the difference metric between them (such as time difference, distance, etc.), is the maximum acceptable difference; ranges from [0, 1], and the closer to 1, the higher the consistency. When When it is considered inconsistent, processing is required.

[0049] The detailed implementation method of the traffic feature extraction module 50 is as follows: The traffic feature extraction module 50 first calculates the travel frequency of the vehicle and counts the number of trips within different time windows. Specifically, it includes: the number of trips within 24 hours , the average number of trips on weekdays , the average number of trips on weekends , the number of trips during peak hours (7:00 - 9:00 and 17:00 - 19:00) and so on. Professional operating vehicles usually have a relatively high travel frequency, especially during peak hours.

[0050] Then, analyze the trip coherence and evaluate the spatio-temporal logical rationality between consecutive trips. Calculate the spatio-temporal connection coefficient of consecutive trips : , where is the end point of the th trip and the start point of the th trip The distance between them (unit: meter), is the reasonable maximum driving speed (unit: meter per second, for example, 60 km / h is equal to 16.67 m / s), and are respectively the start time of the th trip and the end time of the th trip (unit: second), is the total number of trips; The ideal range of the

[0051] value is [0, 1]. The closer it is to 1, the more reasonable the coherence is. If it is too small, it may indicate missed reported trips. If it is too large, it may indicate speeding. , where is the set of parking points, and each point contains longitude and latitude coordinate information; is the clustering radius parameter (usually set to 100 meters), indicating the range of the neighborhood; MinPts is the minimum number of points parameter (usually set to 3), indicating the minimum number of neighborhoods required to form a core point; Cluster is the clustering result, including multiple clusters, and each cluster corresponds to a regular parking area. Calculate the parking duration distribution for each cluster to obtain the average parking duration , the longest parking duration and features such as parking frequency Freq.

[0052] The traffic feature extraction module 50 also calculates the Gini coefficient of the travel trajectory to quantify the inequality of travel distances: , where is the distance of the th trip (unit: meter), is the distance of the th trip (unit: meter), is the total number of trips, is the average travel distance (unit: meter), calculated as is the Gini coefficient, with a value range of [0, 1]. The larger the value, the more uneven the travel distance distribution. The double summation calculates the sum of the absolute differences between all pairs of travel distances and divides by for normalization. The value of professional operating vehicles is usually low because their travel distance distribution is relatively uniform. Preferably, the Gini coefficient threshold is set to 0.3, and vehicles below this threshold are considered to have professional operating characteristics. In addition, the traffic feature extraction module 50 divides the urban area into several sectors. Based on the city center or major transportation hubs, the city is divided into 8 - 12 sectors. Using the boundary checkpoints of each sector to construct a spatio-temporal feature set, including the time distribution, frequency, etc. of entering and leaving the sector. Cluster the boundary checkpoints according to feature similarity and calculate the average residence time of checkpoints within the cluster : , where is the residence duration (unit: minute) of vehicles with a residence time of , is the number of vehicles with a residence time of , is the number of different residence duration types, is the cluster number; is the average residence time (unit: minute) within cluster . The numerator represents the total residence time of all vehicles, and the denominator represents the total number of vehicles. Dividing the two gives the average residence time. Based on the residence time distribution within the cluster, calculate the probability of illegal operation : , is the illegal operation probability index within cluster . The larger the value, the more abnormal it is. Where is the vehicle in cluster The residence time within (unit: minute), is the cluster The average residence time within (unit: minute), and the calculation formula is , where is the cluster The total number of vehicles within; represents the vehicle within the cluster The illegal operation probability index within. The larger the value, the more abnormal it indicates. This index is essentially the absolute value of the Z-score of the residence time, measuring the deviation degree of the vehicle residence time from the average level.

[0053] The present invention further includes a passenger APP feature extraction module 80 and a blockchain evidence storage module 90. The passenger APP feature extraction module 80 is connected to the data interaction module 10 and is used to extract the user rating pattern, search behavior, and payment record from the passenger APP operation behavior data and calculate the illegal operation possibility. Specifically, this module analyzes the rating distribution of users for drivers. Abnormally high or consistent ratings may imply abnormal operation behaviors; analyzes the search behavior of users, including search frequency, content, and time distribution. Frequent searches for specific routes or searches at fixed times may indicate professional operation characteristics; statistically analyzes the amount distribution, frequency, and time pattern of payment records to discover abnormal payment patterns. Based on these features, calculate the illegal operation possibility score : , where , and are the scores calculated based on the rating pattern, search behavior, and payment record respectively, and are all normalized to the [0, 1] interval; , and are the corresponding weights, satisfying , usually taking , indicating the relative importance of the three features. In addition to fusing the spatio-temporal joint feature vector, travel frequency, and trip coherence feature, the data fusion module 60 also fuses the illegal operation possibility score calculated by the passenger APP feature extraction module 80 to form a more comprehensive feature representation and improve the accuracy of judgment.

[0054] The blockchain evidence storage module 90 is connected to the intelligent decision-making module 70 and is used to receive and encrypt and store the identification results of illegal operation vehicles, including information such as MAC address, usage time, license plate, and driving route. This module adopts the distributed ledger technology to ensure that the stored evidence cannot be tampered with and provides reliable technical support for subsequent law enforcement.

[0055] Specifically, the blockchain evidence storage module 90 generates and stores blocks according to a preset time period (such as every 24 hours), and each block contains evidence of illegal operation within a certain time period. Blocks are linked through hashes, and the hash value of each block is calculated as follows: , where, is the hash value of the previous block, is the data content of the current block, is the timestamp, || represents the string concatenation operation, represents the hash function (such as SHA-256); is the hash value of the current block, serving as the unique identifier of the block. This chained structure ensures the immutability of historical data. Any modification to historical blocks will cause the hash values of all subsequent blocks to change, facilitating detection.

[0056] Meanwhile, this module provides an evidence query interface, supporting law enforcement departments to query relevant evidence according to conditions such as license plate number and time period, facilitating evidence collection and law enforcement. The query result contains a complete evidence chain and provides a verification mechanism to ensure the authenticity and integrity of the evidence.

[0057] The method for identifying illegal operation vehicles based on spatio-temporal graph neural network of the present invention includes the following steps: Connect to the mobile payment management platform, urban traffic control platform, and blockchain evidence storage platform through the data interaction module 10 to obtain bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data. This step establishes a multi-source data acquisition channel, providing comprehensive data support for subsequent processing.

[0058] Unify the formats of the bayonet capture data, mobile payment platform trip data, urban traffic control platform data, and passenger APP operation behavior data through the data preprocessing module 20, and perform encoding standardization on the data of each dimension. This step ensures the consistency and comparability of data from different sources, laying a foundation for subsequent analysis.

[0059] Construct a spatio-temporal dynamic relationship graph containing driver nodes, bayonet nodes, vehicle nodes, vehicle type nodes, trip nodes, and parking point nodes based on the encoded and standardized data through the spatio-temporal graph construction module 30. This step organizes various entities and their relationships into a graph structure, facilitating the capture of complex spatio-temporal association patterns.

[0060] Extract spatio-temporal joint feature vectors from the spatio-temporal dynamic relationship graph through the Topology-Enhanced Hierarchical Spatio-Temporal Attention Network 40. This step is the core innovation of the method, and high-quality feature extraction is achieved through a three-layer progressive structure (Topological Space Mapping Layer 41, Probabilistic Attention Weighting Layer 42, and Spatio-Temporal Feature Fusion Layer 43).

[0061] Through the traffic feature extraction module 50, extract the travel frequency and trip coherence features based on the time-space trajectories of bayonet capture data, and calculate the probability of illegal operation. This step analyzes vehicle features from the perspective of traffic behavior and identifies the typical patterns of professional operating vehicles.

[0062] Through the data fusion module 60, fuse the spatio-temporal joint feature vectors, travel frequency, and trip coherence features to extract candidate illegal operating vehicles and form an illegal operating vehicle library. This step integrates multi-dimensional features to improve the comprehensiveness and accuracy of judgment.

[0063] Through the intelligent decision-making module 70, calculate the weight coefficients of each dimension feature, screen the vehicles in the illegal operating vehicle library based on the entropy value method, and determine the illegal operating vehicles. This step realizes the final determination of illegal operation and outputs the recognition result.

[0064] The entire method process forms a complete closed-loop for identifying illegal operating vehicles. From data acquisition, preprocessing, feature extraction to final decision-making, each link is closely connected and works together to achieve efficient and accurate identification of illegal operating vehicles.

[0065] The illegal operating vehicle identification system based on spatio-temporal graph neural network of the present invention can be deployed on the server cluster of the urban traffic management center and connected to each data platform through a standard interface. The system supports distributed deployment and can be flexibly expanded according to the city scale and data volume.

[0066] In a practical application case, this system is deployed on the traffic management platform of a first-tier city, processing about 5 million bayonet capture records and 3 million trip data every day. The system analyzes about 1 million active vehicles and identifies about 200 suspected illegal operating vehicles on average every day. After verification by the traffic law enforcement department, the recognition accuracy rate reaches more than 85%, greatly improving the law enforcement efficiency. At the same time, the blockchain evidence storage function of the system provides reliable evidence for law enforcement, and the law enforcement success rate is increased by about 40%.

[0067] In addition, the system also has strong adaptability and can automatically adjust parameters according to the characteristics of different cities. For example, in tourist cities, the system will appropriately lower the judgment threshold around scenic spots to reduce misjudgment of tourist vehicles; in cities with serious traffic congestion, the system will adjust the trip coherence evaluation criteria to adapt to the special traffic environment.

[0068] In summary, the illegal operation vehicle identification system and method based on the spatio-temporal graph neural network provided by the present invention achieve high-precision identification of illegal operation vehicles through an innovative topological enhanced hierarchical spatio-temporal attention network, and have broad application prospects and significant social value.

[0069] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. The illegal vehicle identification system based on spatiotemporal graph neural network is characterized by: include: The data interaction module is used to connect the mobile payment management platform, the urban traffic control platform and the blockchain evidence storage platform to obtain the card entrance capture data, the mobile payment platform travel data, the urban traffic control platform data and the passenger APP operation behavior data; A data preprocessing module, connected to the data interaction module, is used to unify the format of the card-snatching data, the mobile payment platform travel data, the urban traffic control platform data and the passenger APP operation behavior data, and to encode and standardize the data of each dimension; A spatiotemporal graph construction module, connected to the data preprocessing module, for constructing a spatiotemporal dynamic relationship graph containing driver nodes, checkpoint nodes, vehicle nodes, vehicle model nodes, trip nodes and parking point nodes based on the coded and standardized data; A topologically enhanced hierarchical spatiotemporal attention network, connected to the spatiotemporal graph building module, comprising a topological space mapping layer, a probabilistic attention weighting layer and a spatiotemporal feature fusion layer, for extracting a spatiotemporal joint feature vector from the spatiotemporal dynamic relationship graph; A traffic feature extraction module, connected to the data preprocessing module, is used to extract travel frequency and trip continuity features based on the time-space trajectory of the camera capture data, and calculate the probability of illegal operation; A data fusion module, connected to the topologically enhanced hierarchical spatiotemporal attention network and the traffic feature extraction module, for fusing the spatiotemporal joint feature vector, the travel frequency and the trip continuity feature, extracting candidate illegal operating vehicles, and forming an illegal operating vehicle database; The intelligent decision-making module is connected to the data fusion module and is used to calculate the weight coefficient of each dimensional feature, screen the vehicles in the illegal operation vehicle database based on the entropy method, and determine the illegal operation vehicles.

2. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The topological space mapping layer is used to: The urban road network is represented as a graph structure, where intersections are nodes and roads are edges; Assigning a weight to each edge of the graph structure, wherein the weight represents a road travel time or a road travel difficulty; Construct a road adjacency matrix to represent the connectivity of the road network; Map the vehicle GPS track point sequence to the nearest road edge; Compute the path representation of the trajectory in the topological space; Extract the topologically invariant features of the path; Calculate the topological persistence index of the trajectory, which represents the trajectory structure stability; Identify topological features in trajectories; Generate trajectory topology signatures.

3. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The probabilistic attention weighted layer is used to: Decompose the trajectory into sub-trajectories of different time scales; Extract features for each time scale to capture behavioral patterns at different time granularities; Construct a multi-scale feature matrix to represent the feature distribution at different time scales; Construct a conditional probability model to quantify the importance of features in determining illegal operations; Calculate the information gain of each feature as the basis for attention weight; Design three attention heads: time, space, and spatiotemporal joint attention heads; The attention mechanism is used to highlight abnormal behavior characteristics and improve detection sensitivity.

4. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The spatiotemporal feature fusion layer is used to: Analyze the complementarity and redundancy of characteristics from different sources; Remove highly correlated features and retain complementary information; Construct feature dependency graph to guide feature fusion strategy; Adjust the fusion weights based on city characteristics, time periods, and weather factors; Construct joint features representing the starting area, travel path, and passing areas; Preserve high-order correlations between features; Generate the final spatiotemporal joint feature vector.

5. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The data preprocessing module is also used for: Standardize time information and unify the time representation format of data from different sources; Standardize spatial information and map location information to a unified geographic coordinate system; Standardize vehicle information and unify vehicle identification representations from different sources; Perform outlier detection and processing to identify and process data that significantly deviates from the normal range; Perform missing value completion and make reasonable inferences about missing data based on spatiotemporal correlation; Perform data consistency verification to check whether the same information from different sources is consistent.

6. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The spatiotemporal graph construction module is also used for: Set driving habits and driving time attributes for the driver node; Set geographic location and passing frequency attributes for the checkpoint node; Set the vehicle model and license plate color attributes for the vehicle node; Set the function type and regional characteristic attributes for the location node; Construct vehicle-checkpoint edges to represent the time and frequency of vehicles passing through the checkpoints; Construct vehicle-location edges to represent the vehicle’s stay time and frequency at the location; Construct driver-vehicle edges to represent driving association strength and time distribution; Construct location-location edges to represent the travel frequency and travel time between locations; Implement dynamic graph updates and update the graph structure when new data arrives.

7. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The traffic feature extraction module is also used for: Calculate the travel frequency and count the number of trips in different time windows; Analyze the continuity of the trip and evaluate the rationality of the time and space logic between consecutive trips; Extract parking patterns, identify regular parking spots and duration distribution; Calculate the Gini coefficient of travel trajectories to quantify the inequality of travel distances; Setting a Gini coefficient threshold to determine professional driver status; Divide the urban area into several sectors; Use the boundary checkpoints of each sector to construct a spatiotemporal feature set; According to the feature similarity of the feature set, the boundary checkpoints are clustered into several clusters; Calculate the average dwell time at checkpoints within the cluster and the probability of illegal operation.

8. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The system further comprises: A passenger APP feature extraction module, connected to the data interaction module, is used to extract user rating patterns, search behaviors and payment records from the passenger APP operation behavior data and calculate the possibility of illegal operation; The data fusion module is also used to fuse the possibility of illegal operation.

9. The illegal vehicle identification system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The system further comprises: A blockchain evidence storage module, connected to the intelligent decision-making module, is used to receive and encrypt and store the identification results of the illegally operated vehicle, including MAC address, usage time, license plate, and driving route information; Block storage is performed according to a preset time period; Providing verifiable evidence for law enforcement.

10. An illegal vehicle identification method based on spatiotemporal graph neural network, using the system according to any one of claims 1 to 9, characterized in that: include: Through the data interaction module, the mobile payment management platform, the urban traffic control platform and the blockchain evidence storage platform are connected to obtain the card gate capture data, the mobile payment platform travel data, the urban traffic control platform data and the passenger APP operation behavior data; The data preprocessing module unifies the format of the card entrance capture data, the mobile payment platform travel data, the urban traffic control platform data and the passenger APP operation behavior data, and encodes and standardizes the data of each dimension; A spatiotemporal dynamic relationship graph including driver nodes, checkpoint nodes, vehicle nodes, vehicle model nodes, trip nodes and parking point nodes is constructed based on the encoded and standardized data through a spatiotemporal graph construction module; Extracting a spatiotemporal joint feature vector from the spatiotemporal dynamic relationship graph through a topologically enhanced hierarchical spatiotemporal attention network, wherein the topologically enhanced hierarchical spatiotemporal attention network includes a topological space mapping layer, a probabilistic attention weighting layer, and a spatiotemporal feature fusion layer; The traffic feature extraction module extracts the travel frequency and trip continuity features based on the time-space trajectory of the camera capture data, and calculates the probability of illegal operation; The spatiotemporal joint feature vector, the travel frequency and the trip continuity feature are fused through a data fusion module to extract candidate illegal operating vehicles and form an illegal operating vehicle database; The weight coefficient of each dimensional feature is calculated through the intelligent decision-making module, and the vehicles in the illegal operation vehicle database are screened based on the entropy method to determine the illegal operation vehicles.

Images